1. Computational and Systems Biology
  2. Epidemiology and Global Health

Linking glycemic dysregulation in diabetes to symptoms, comorbidities, and genetics through EHR data mining

  1. Isa Kristina Kirk
  2. Christian Simon
  3. Karina Banasik
  4. Peter Christoffer Holm
  5. Amalie Dahl Haue
  6. Peter Bjødstrup Jensen
  7. Lars Juhl Jensen
  8. Cristina Leal Rodríguez
  9. Mette Krogh Pedersen
  10. Robert Eriksson
  11. Henrik Ullits Andersen
  12. Thomas Almdal
  13. Jette Bork-Jensen
  14. Niels Grarup
  15. Knut Borch-Johnsen
  16. Oluf Pedersen
  17. Flemming Pociot
  18. Torben Hansen
  19. Regine Bergholdt
  20. Peter Rossing  Is a corresponding author
  21. Søren Brunak  Is a corresponding author
  1. University of Copenhagen, Denmark
  2. Odense University Hospital, Denmark
  3. Steno Diabetes Center Copenhagen, Denmark
  4. Rigshospitalet, Denmark
  5. Holbæk Hospital, Denmark
  6. Herlev-Gentofte Hospital, Denmark
  7. Technical University of Denmark, Denmark
https://doi.org/10.7554/eLife.44941
Share this article
Cite this article
  1. Isa Kristina Kirk
  2. Christian Simon
  3. Karina Banasik
  4. Peter Christoffer Holm
  5. Amalie Dahl Haue
  6. Peter Bjødstrup Jensen
  7. Lars Juhl Jensen
  8. Cristina Leal Rodríguez
  9. Mette Krogh Pedersen
  10. Robert Eriksson
  11. Henrik Ullits Andersen
  12. Thomas Almdal
  13. Jette Bork-Jensen
  14. Niels Grarup
  15. Knut Borch-Johnsen
  16. Oluf Pedersen
  17. Flemming Pociot
  18. Torben Hansen
  19. Regine Bergholdt
  20. Peter Rossing
  21. Søren Brunak
(2019)
Linking glycemic dysregulation in diabetes to symptoms, comorbidities, and genetics through EHR data mining
eLife 8:e44941.
https://doi.org/10.7554/eLife.44941
  2. Download BibTeX
  3. Download .RIS

Abstract

Diabetes is a diverse and complex disease, with considerable variation in phenotypic manifestation and severity. This variation hampers the study of etiological differences and reduces the statistical power of analyses of associations to genetics, treatment outcomes, and complications. We address these issues through deep, fine-grained phenotypic stratification of a diabetes cohort. Text mining the electronic health records of 14,017 patients, we matched two controlled vocabularies (ICD-10 and a custom vocabulary developed at the clinical center Steno Diabetes Center Copenhagen) to clinical narratives spanning a 19 year period. The two matched vocabularies comprise over 20,000 medical terms describing symptoms, other diagnoses, and lifestyle factors. The cohort is genetically homogeneous (Caucasian diabetes patients from Denmark) so the resulting stratification is not driven by ethnic differences, but rather by inherently dissimilar progression patterns and lifestyle related risk factors. Using unsupervised Markov clustering, we defined 71 clusters of at least 50 individuals within the diabetes spectrum. The clusters display both distinct and shared longitudinal glycemic dysregulation patterns, temporal co-occurrences of comorbidities, and associations to single nucleotide polymorphisms in or near genes relevant for diabetes comorbidities.

Introduction

Electronic Health Records (EHRs) contain patient characteristics from different data layers including text narratives, assigned diagnosis codes, biochemical values, and prescription data. These data types display a high degree of complementarity, providing an excellent basis for deep phenotyping and patient stratification. Recent studies have shown how structured data derived from EHRs can be used to assess phenotypic variability of different disease areas (Li et al., 2015; Dahlem et al., 2015; Doshi-Velez et al., 2014; Kho et al., 2011; Kho et al., 2012). While the use of structured EHR data in many instances resembles traditional registry- or biobank-based research, the inclusion of unstructured data such as clinical narratives allows for the definition of even more fine-grained phenotypes, which could lead to novel subgroup stratifications (Li et al., 2015; Roque et al., 2011; Miotto et al., 2016).

A vast amount of information on symptoms, lifestyle, complications, and comorbidities is available from clinical narratives in unstructured EHR data. Text mining applying natural language processing (NLP) algorithms is one strategy, but simpler approaches have also been shown to be valuable in the context of clinical text, for reviews see Jensen et al. (2012) and Denny (2012). These methods work across language barriers and have been successfully implemented in for example adverse drug reaction detection (Warrer et al., 2012), subgrouping of chronic obstructive pulmonary disease (Fu et al., 2015), cancer subgrouping (Chen et al., 2015), and classification of epileptic children (Pereira et al., 2013). Such studies show the possibilities of using and integrating different parts of EHRs for matching phenotypically similar subgroups to biomarker data, which is key to improved treatment and characterizing etiological differences.

Several large initiatives have been established for utilizing EHRs, including the Electronic Medical Records and Genomics (eMERGE) consortium of DNA biorepositories that links genetic data with electronic medical records (McCarty et al., 2011; Gottesman et al., 2013), and EMR-driven nonnegative restricted Boltzmann machines (eNRBM) which use unsupervised learning for analyzing EHRs (Tran et al., 2015). Furthermore, other studies have used general approaches for finding direct and inverse comorbidities (Doshi-Velez et al., 2014; Roque et al., 2011; Gligorijevic et al., 2016).

Diabetes Mellitus (DM) is a difficult disease to stratify (American Diabetes Association, 2017). DM covers etiologically different metabolic disorders that exhibit the same phenotype, hyperglycemia, due to either insufficient insulin production relative to insulin demand or insulin resistance. Although DM is classified into different major subtypes, it has been hypothesized to represent a disease continuum rather than strict distinct disease subtypes (American Diabetes Association, 2017; Flannick et al., 2016). One recent data-driven study used five subgroups of adult-onset diabetes and clustered six parameters from the structured data of the EHR (Ahlqvist et al., 2018). DM is a complex disorder associated with several comorbidities and organ complications. These can be classified as macrovascular complications that is cardiovascular disease, and microvascular complications resulting in eye, kidney, and nerve damage. Cardiovascular complications alone are responsible for 50–80% of all-cause mortality in diabetes patients (Laakso, 2001). The severity of complications is affected by glycemic dysregulation, that is increased or fluctuating blood glucose levels (Stratton et al., 2000; UK Prospective Diabetes Study Group, 1998a; UK Prospective Diabetes Study Group, 1998b; Nathan et al., 1993), and successful reduction and prevention of diabetic complications have been observed when the glycemic dysregulation is reduced or removed (Stratton et al., 2000; UK Prospective Diabetes Study Group, 1998a). Therefore, risk factors for glycemic dysregulation are crucial to diabetes progression (Ahlqvist et al., 2015). Known risk factors for complications include age, diabetes duration, polypharmacy, comorbidities (Juarez et al., 2012), increased levels of circulating triglyceride and LDL-cholesterol, and lower levels of HDL-cholesterol (Saudek et al., 2006; Giannini et al., 2011; Bitzur et al., 2009). Finding new risk factors that can help classify poorly regulated versus well-regulated diabetes, such as other biochemical variables or genetic variants, could improve treatment and reduce diabetic complications.

In this study, we utilized the unstructured data of EHRs and performed a deep phenotypic characterization of a Danish diabetes cohort of 14,017 individuals, aged 18 to 101 at the end of the study, using vocabularies comprising both diagnosis codes and ‘exposome’ related terms. We used text-mined and assigned diagnosis codes to stratify the cohort and described it using both physiological and genetic variation data. The unstructured EHR data enabled us to classify patients based on their level of glycemic dysregulation and to identify potential biochemical and genetic markers associated with dysglycemia.

Results

Text mining the EHR corpus

The general aim of the text mining effort was to obtain a richer phenotypic characterization of each patient. Initially, each patient had in 4.9 assigned codes on average. Applying text mining with two vocabularies (ICD-10 and SDC-custom) resulted in a 4-fold increase to 18.6 codes per patient. Moreover, the distribution of codes across ICD-10 chapters changed considerably when adding the text-mined codes, with chapters I, VII, XVIII and XIX showing the largest increases (6, 15, 25 and 22-fold increase, respectively) (Figure 1). This illustrates the difference between the assigned diagnosis codes from the structured data and the much more symptom-rich codes detected by text mining.

Figure 1 with 6 supplements see all
Download asset Open asset
Comparison of distributions of ICD-10 diagnosis codes with and without text mining.

(A) Percentage of diagnosis codes belonging to the different ICD-10 chapters and the relative increase in diagnosis codes from the different chapters when combining the text-mined and assigned codes. (B) Age distributions of text-mined and assigned ICD-10 diagnosis codes from the SDCC corpus divided into the 21 ICD-10 chapters.

Figure 1—source data 1

Diagnosis code breakdown data.

https://cdn.elifesciences.org/articles/44941/elife-44941-fig1-data1-v1.txt
Figure 1—source data 2

Age distribution data.

https://cdn.elifesciences.org/articles/44941/elife-44941-fig1-data2-v1.txt

Comorbidity clustering based on text-mined and assigned diagnosis codes

For each patient, the assigned and text-mined ICD-10 codes were combined to create a patient-specific diagnosis-vector where the primary diabetes type (E10 or E11) was not included. Contrary to cancer for example, where the ICD-10 diagnoses are quite reliable and highly detailed, the primary codes in a multi-organ disease like diabetes are used in a fuzzier way, as the knowledge on robust diabetes subtypes and their characteristics in the context of comorbidities is quite limited. We do therefore not want the clustering to be driven by the broad, less etiology-relevant primary codes from the endocrinology chapter, but rather by more objectively observed symptoms, other diseases and lifestyle features. Following code-abundance normalization and BM-25 correction the vectors were clustered using MCL producing 172 clusters (mean = 65 patients, min = 11, max = 979, median = 40), in which 11,208 patients (80.47%) were included Figure 2A. The remaining 2720 patients (19.53%) were in clusters with ten or less patients and were therefore omitted from subsequent analyses.

Figure 2 with 5 supplements see all
Download asset Open asset
Phenotypic clusters found in the SDCC cohort.

The clustering was created with diagnosis vectors of 13,928 patients (with text in the record) comprising both text-mined and assigned ICD-10 codes. A total of 172 clusters were created, where 11,208 patients (80.47%) were captured in the clustering (clusters with five or less patients were discarded for statistical reasons). (A) Each node represents a patient within the corpus colored by the association to one of the 172 unique clusters. (B) The 71 clusters with at least 50 patients colored with the same palette as in (A).

Even though codes for the primary diabetes type were not part of the diagnosis vectors, specific clusters were significantly enriched for T1D patients (cluster 1: N = 506, adj. p-value=9.3e-51 and cluster 9: N = 101 adj. p-value=1.2e-10). Other clusters had significantly more T2D patients than expected (cluster 3: N = 233, adj. p-value=9.1e-10, cluster 5: N = 170, adj. p-value=3.8e-13 and cluster 6, N = 158 adj. p-value=8.4e-17). In addition, we observed a cluster significantly enriched with the ICD-10 term E13: other diabetes (cluster 25, N = 93, adj. p-value=1.8e-142), which includes diabetes due to genetic defects, post-pancreatectomy diabetes and post-procedural diabetes. Several other clusters had a mix of T1D and T2D patients according to the assigned codes. Further characteristics of the laboratory data and prescription data as well as the clusters regarding sex, age, observational time, years with diabetes etc. can be found in Supplementary files 1–3 and in Figure 1—figure supplement 1, Figure 1—figure supplement 2, Figure 1—figure supplement 3, Figure 1—figure supplement 4, Figure 2—figure supplement 1, Figure 2—figure supplement 2, Figure 2—figure supplement 3, Figure 2—figure supplement 4. The robustness of the clustering was found to be high (see description in Materials and methods and Figure 2—figure supplement 5). To maintain power in subsequent analyses we focused on clusters with at least 50 patients (71 clusters comprising 8652 patients, Figure 2B).

Enriched comorbidity and symptom patterns in diabetes patient clusters

The 71 clusters (Figure 2B) were grouped by hierarchical clustering, using distances obtained from cluster specific symptoms from the ICD-10 chapter XVIII (level 1). Six main groups and an outlier (cluster 70) were found containing 5, 8, 21, 11, 7 and 18 of the original clusters, respectively. The symptom groups are illustrated by the branch colors in Figure 3. The nodes represent the 71 clusters each depicted as a pie chart displaying the comorbidities and symptoms that are significantly enriched (adj. p-value≤0.05), see Supplementary file 4 for details on the enrichment and p-values.

Figure 3
Download asset Open asset
Hierarchical clustering based on enriched comorbid ICD-10 diagnoses.

The comorbidities present in a minimum of 10 patients and significantly enriched (adj. p-value<=0.05) in each cluster are shown in the pie charts. The number of significant codes ranges from 1 to 10. Each color corresponds to an ICD-10 code chapter as listed in the legend of Figure 1. Six main groups and an outlier (cluster 70) resulted, and the colors of the dendrogram branches indicate to which hierarchical groups the clusters belong. The size of the pie charts represents the average diabetes duration (years with diabetes) divided into six bins. The 21 clusters where at least 50% of the patients have three or more HbA1c severity parameters are marked with a red line surrounding the pie chart.

The 71 clusters were defined based on the associated comorbidities, excluding DM without complications, and from the pie charts we observed that distinct diagnoses do indeed characterize the clusters. For example, ICD-10 code N40: Benign prostatic hyperplasia for cluster 56, L40: Psoriasis for cluster 16, F20: Schizophrenia for cluster 47, K29: Functional intestinal disorders for cluster 17, and Z94: Transplanted organ and tissue status for cluster 42. Using Fisher’s exact test, we found that: Symptoms related to skin and subcutaneous tissue (adj. p-value<0.001) characterized symptom group five and Symptoms related to digestive system and abdomen; cognition, perception, emotional state and behavior; and general symptoms and signs (adj. p-value<0.001 for all) characterized symptom group 3. These results correspond well to the enriched codes observed in Figure 3, as was the case for the other enriched codes across the 71 clusters within the six symptom groups.

Genomic characterization by SNP association of phenotypically determined clusters

We evaluated the 71 clusters in the six symptom groups, plus the outlier cluster, for SNPs that could characterize the different groups (details on the genetic data can be found in the ‘Genomic characterization’ section under Materials and methods). The five highest association signals (independent) for each group are shown in Supplementary file 5. Only results from analyses with more than 15 cases and a well-calibrated QQ-plot (visual inspection and a lambda inflation factor >0.96) are reported. Accordingly, clusters 1–5, 7–9, 12, 15–18, 21–23, 26, 31, 35, 39, 45, 46, and 66, as well as all aggregated symptom groups, met the criteria. The median coverage of the symptom clusters was 31% [range: 10–67]. SNPs characterizing the symptom groups were found in several instances and association signals to disease-associated genes were also found for several of the clusters (Figure 3). Most frequently found, unsurprisingly, were genes associated by GWAS to diabetes or diabetes-related cardio-metabolic traits (cluster 3: MYO3B, cluster 4: DAPK1, cluster 5: LPIN2, cluster 7: SAMD4A and FHIT, cluster 8: ERG and PLCB1, cluster 12: MYT1L, cluster 15: UBE2WP1, cluster 16: ADARB2, CDKAL1, and CLIP1, cluster 17: C8orf37-AS1, cluster 21: FHOD3 and MCF2L, cluster 24: MTCL1, cluster 26: NTM, cluster 31: PCDH15, CDH4, and DCTD, cluster 31: KLF12, cluster 39: FHOD3, cluster 45: IGF1R, BCAS3, and TENM4, cluster 46: NRXN3). Cluster eight is characterized by cardiovascular complications, and three of the top ranking genes for this cluster have been associated with LDL peak particle diameter (THBS4; Rudkowska et al., 2015), abdominal aortic aneurysm (ERG; Jones et al., 2017), pulse pressure (ERG; Warren et al., 2017), and diastolic blood pressure (PLCB1; Warren et al., 2017). Cluster 21 is enriched for the ICD-10 diagnosis foot ulcer (L97), and MCF2L, one of the top ranking genes for cluster, has been associated with both end-stage coagulation (Williams et al., 2013) and prothrombin time (Tang et al., 2012). In total, of the top five association signals that were mapped to genes (n = 103) we found five (CDKAL1, DCDC2C, KLF12, LPIN2, TLE1) to be related with diabetes.

Comorbidity pairs and patterns within symptom related clusters

We detected codes occurring significantly more or less together within and across the symptom groups (Fischer’s test with Bonferroni adjusted p-values<=0.01) defining distinct comorbidity pairs. If the comorbidity pairs covered more than 100 unique codes (symptom groups 4 and 7) we extracted only the most significant pairs until these pairs consisted of 100 unique codes.

Figure 4A illustrates the comorbidity correlations for the six main symptom groups where each pairwise interaction has a comorbidity score (see Material and methods). To characterize whether a diagnosis occurred significantly more before or after another, we made this analysis in a temporal manner. Figure 4B illustrates the comparison of the first diagnosis (row) to the second diagnosis (column). We found that especially the diagnoses related to diabetes (E13, O24), diabetes with complications (shortened to E10 and E11), obesity (E66), diseases of the pancreas (K86), poly- and proteinuria (R35 and R80), and to some extent hypertension and ischemic heart disease (I10, I20, I21, I25) are observed before other diagnoses (blue indicates that the row diagnosis is observed prior to the column diagnosis more than expected, and red indicates the opposite). Focusing on the different symptom groups, we detected which comorbidity pairs were unique in the different groups, and Figure 4C displays these unique comorbidity interactions.

Figure 4
Download asset Open asset
Comorbidity patterns within the six symptom groups.

(A) Comorbidity correlations between the combined symptom groups. (B) Asymmetric comorbidity matrix for observing row diagnosis codes before column diagnoses. First, we calculated Bonferroni corrected p-values for diagnosis pair directionality, second, we extracted the top 100 unique diagnosis codes pairs with lowest adjusted p-values and lastly, we calculated a comorbidity score (CS) by using the log2 of observing the pair more or less than expected. The heat-map colors reflect the CS quantification. (C) Comorbidity pairs unique for each of the symptom groups. All interactions are observed significantly more (blue) or less (red) than expected (adj. p-value<=0.01). Arrows indicate that the diagnoses are observed in the particular order (Fischer’s exact test with Bonferroni correction p-value<=0.01). Node size indicates in how many symptom groups the diagnosis code is observed in, ranging from one group (the diagnosis is unique for the group, largest nodes) to six groups (all groups have the code, smallest nodes).

Figure 4—source data 1

Comorbidity pattern data.

https://cdn.elifesciences.org/articles/44941/elife-44941-fig4-data1-v1.txt

In symptom group two we found that L84: corns and callosities is observed significantly more together within patients with T1D than T2D (CS = 1.24, adj. p-value=4.06e-15 and CS = −1.58, adj. p-value=1.25e-03, respectively). Temporal analysis of diagnosis occurrence showed that T1D is observed before L84 (Figure 4B, mean time difference = 8.3 years, adj. p-value=1.01e-39). Corns or callosities are unproblematic in healthy people, but in diabetes patients they can cause skin defects that increase the risk for additional complications, for example foot ulcers which can lead to amputations (Apelqvist et al., 2000; Hunt, 2011).

Although not observed significantly together within any clusters the temporal analysis showed that the time between T2D and elevated blood glucose levels (R73) is significantly shorter in symptom group two than in groups 4, 5 and 6 (mean time = 2.2 days; adj. p-value=6.45e-04, 3.29e-06 and 2.73e-06, respectively).

In symptom group 5, five of the eleven clusters are enriched with ICD-10 codes from chapter XIII: Diseases of the musculoskeletal system and connective tissue, especially dorsopathies, spondylopathies and soft tissue disorders. Further, these diagnoses are observed exclusively in this group and show unique disease co-occurrence patterns, for example M48-M54 (other spondylopathies and dorsalgia, CS = 1.01, adj. p-value=1.8e-04) and M43-M47 (deforming dorsopathies and spondylosis, CS = 1.54, adj. p-value=1.91e-06). One of the top ranked genetic associations for this cluster (rs76548985, p-value=1.43e-06) is LINC00351, associated with sporadic amyotrophic lateral sclerosis (Xie et al., 2014). It is worth noting that clusters 8, 22, 33, 35, 45 within symptom group five are all enriched for drugs from ATC chapter A10B: blood glucose lowering drugs, excluding insulin (adj. p-value<0.05), and all but cluster eight are associated with glycemic dysregulation.

Within symptom group 7, we observed two diagnosis pairs less than expected: E11-E13 (CS = −1.46, adj. p-value=1.57e-04), and K86-E78 (CS = −1.24, adj. p-value=5.46e-04). Hence, this group contains patients where T2D and other diabetes as well as diseases of the pancreas and disorders of lipoprotein metabolism are not given together. In contrast, I34: nonrheumatic mitral valve disorder is observed more often than expected together with heart failure (I50, CS = 1.83, adj. p-value=0.009) and atrioventricular and left bundle-branch block (I44, CS = 1.53, adj. p-value=0.0018). Interestingly, one of the top genetic signals for symptom group seven maps to MIR8052 (rs6590490, p-value=3.14e-07) that has been associated with pulse pressure (Warren et al., 2017). Comparably, among the top genetic signals for symptom group 4, a group where are large proportion of the patients are characterized by hypotension (I95) and vertigo (R42), are ANLN that has been associated with systolic blood pressure (Parmar et al., 2016).

Glycemic dysregulation

We evaluated five different parameters associated with glycemic dysregulation (glycemic dysregulation, hyperglycemia, check-point detection of fluctuating HbA1c levels, HbA1c level at diabetes onset and amount of HbA1c observations above diagnosis threshold for T1D and T2D [53 mmol/mol]) and found that 2942 patients did not meet any threshold criterion, 2484 met one, 4647 two, 4057 three, 531 four, and 22 met all five criteria. The distribution of HbA1c measurements for T1D and T2D is shown in Figure 1—figure supplement 5. First, we investigated whether there was any difference in mean values of the 20 different biochemical tests (see Material and methods) and subsequently we applied a Kolmogorov-Smirnov test to assess how these distributions differed. We found that the means of 14 of the different biochemical tests were differently distributed between the six groups (adj. p-value<0.01) of patients with different number of dysregulation parameters, and furthermore observed a distinct difference between the not-or-slightly dysregulated patients (groups 0 to 2) and the middle-or-highly dysregulated patients (groups 3 to 5) (Figure 1—figure supplement 6). The group with five parameters showed no significant difference, due to the low number of patients (N = 22). The group with 3–5 parameters showed higher levels of triglyceride and HbA1c, and lower levels of sodium, urine creatinine, C-peptide, hemoglobin, diastolic blood pressure and height. Elevated levels of HbA1c, triglyceride, LDL-cholesterol and cholesterol and lower levels of HDL-cholesterol are known biochemical values associated with glycemic dysregulation and thus verified our findings.

The detection of higher levels of potassium and plasma creatinine as well as the lowered sodium, urine creatinine, hemoglobin levels indicates that these biochemical tests might be used in future prediction of glycemic dysregulation. Glycemic dysregulation is expected to cause renal problems, (identified by elevated plasma creatinine and elevated urine albumin) and hypertension, which is treated with RAS blocking agents (ACE inhibitors and angiotensin two receptor blockers) and diuretic agents, which elevate potassium and lower sodium. The treatment profile of this patient group revealed an enrichment of patients treated by RAS blocking agents in most of the clusters. Based on these observations we considered having at least three of the parameters as the best approximation for a definition of potential glycemic dysregulation.

Using the results from the biochemical analysis, we divided the cohort in two: those with at least three parameters associated with glycemic dysregulation, and those with two or less. In the 71 clusters defined above, 21 had more than 50% patients with at least three parameters (Figure 3, red circles). We found 10 of the 21 clusters in symptom group 3 of which, cluster 5, 24, and 47, were enriched for poor compliance when using the SDC-custom dictionary (adj. p-value=5.9e-03, 1.9e-03 and 2.6e-02, respectively). By further investigating the enrichment of SDC-custom terms (adj. p-values≤0.05) we found that the majority of the 21 clusters had terms related to cardiovascular complications (e.g. beta blocks, ischemia, diuretics and bypass), kidney complications (e.g. nephropathy, edema and albuminuria), metabolic complications (hypoglycemia and insulin chock) and neurologic related disorders (e.g. neuropathy and loss of memory). Furthermore, all the patients in cluster 47 have schizophrenia (N = 76, adj. p-value=8e-141), and behavioral features might therefore account for the glycemic dysregulation. The same could be the case for cluster 24, in which all have epilepsy (N = 108, adj. p-value=7.6e-186).

Genetic characterization of dysregulated patients

To assess if glycemic dysregulation is a diabetic complication or evidence of disease etiology, we further tested whether any SNPs were associated with glycemic dysregulation (n = 2,120). The five top associating signals map to NCKAP5, CLNK, PSD3, KPNA5, and LINC00333 (Supplementary file 5), although not reaching genome-wide significance. Interestingly, two of the genes associated with schizophrenia (LINC00333 [Goes et al., 2015] and NCKAP5 [Draaken et al., 2015]) and PSD3 have also been associated to traits related to urinary and blood metabolite levels, metabolic traits, and triglyceride levels (Raffler et al., 2015; Teslovich et al., 2010; Shin et al., 2014; Rueedi et al., 2014). However, none of the five top ranked genes have been previously linked to glycemic levels or diabetic dysregulation.

Discussion

Previous studies using EHRs in diabetes research have focused on improving clinical decision making (O'Connor et al., 2011), clinical prediction (Miotto et al., 2016), patient management (Cebul et al., 2011), mortality risk (Pantalone et al., 2009; Pantalone et al., 2010), genetic risk factors (Kho et al., 2012), and subgroup identification (Li et al., 2015). Only the study by Miotto et al. (2016) used the different layers of the EHRs, aimed at predictive measures of clinical outcome. A study from the eMERGE consortium extracted phenotypes from EHR narratives by using NLP-based methods (Kho et al., 2011). They used EHR for phenotypic characterization of five main diseases, but a fine-grained analysis of phenotypic characterization within the diseases was not performed. Further, NLP was included only in the phenotypic determination of three of the diseases, not for diabetes determination.

Stratification and subdivision of diabetic cohorts have typically been performed on homogeneous data sets within specific diabetes types such as T1D, T2D, or gestational diabetes (Perry et al., 2012; Ren et al., 2016; Lin et al., 2012; Achenbach et al., 2004). One of the more recent stratification studies of diabetes patients is Li et al. (2015) that identified subtypes of T2Ds of mixed ethnicity using the structured part of EHRs. They detected three distinct subgroups that could be linked to significant SNPs through gene-disease associations in a patient-unspecific manner. Further, elevated HbA1c levels were used to explain one subgroup with microvascular diabetic complications. In contrast to the study by Li et al., we have taken the stratification and characterization several steps further both by investigating a heterogeneous diabetic cohort almost five times as large and obtaining the full comorbidity pattern and symptoms relatedness through mining of the text-narratives using both an ‘exposure-oriented’ and a diagnosis-based dictionary. In addition, we used the biochemical data to produce a severity classification (the five parameters of glycemic dysregulation) and integrating this with both the text-mined and assigned diagnoses, we were able to determine many different, more homogeneous groups of patients with shared symptoms and comorbidities, as well as different levels for glycemic dysregulation.

Another recent diabetes stratification study by Ahlqvist et al. (2018) used a data-driven approach and k-means clustering to subgroup adult-onset diabetes and characterize five subgroups showing differing disease progression and risk of diabetes complications. However, this approach concerned only individuals with type 2 diabetes and a characterization based on six parameters (glutamate decarboxylase antibodies, age at diagnosis, BMI, HbA1c, and homoeostatic model assessment 2 estimates of β-cell function and insulin resistance), and thus clinical narratives, medication, and genetics were not used as we have done in this study.

The text mining approach used in relation to ICD-10 codes was based on level three rather than the more detailed level four since it would increase tremendously the dimensionality of the feature space. While this obviously reflects a less deep phenotyping, for a data set of this size many level four codes would be unique, likely leading to a less stable subsequent clustering and analysis. In fact, our attempt to use the much more fine-grained SNOMED-CT terminology confirmed that a data set needs to be very large for such a fine-grained vocabulary to be useful.

In this work, we deliberately excluded the primary diabetes types without complications, T1D and T2D, and thereby constructed a stratification of the cohort driven solely by comorbidities, complications, other diseases, and symptoms. However, combining different diabetic subtypes can be problematic, since their etiologies differ and disease progression is different across diabetes types, treatment, compliance and lifestyle (Adeghate et al., 2006). Our focus was not to characterize specific comorbidity-related groups within a certain diabetes type, since extensive epidemiological studies of this kind have been done previously. Instead, we focused on the diabetes continuum with the aim of investigating whether it was possible in an unsupervised manner to detect relevant and meaningful diabetic subgroups by comorbidities, symptoms, or level of glycemic dysregulation. Further, we detected novel biochemical and genetic candidates that might relate these to the different cohort subdivisions, such as shared symptom patterns for phenotypically similar patients and the level of glycemic dysregulation. These biochemical and genetic candidates could be potential risk factors for additional complications, especially concerning glycemic dysregulation, that could be verified by further experimental studies. As the cohort is enriched for sicker patients with diabetes melitus complications the features and the overall grouping described would not necessarily be the same if another cohort dominated by prediabetes individuals would have been analyzed.

Despite our focus on the phenotypic variation among diabetes patients, the stratification is restricted by the limited coverage of the genetic data, which lowers the power considerably. We were able to obtain genetic data for 2337 patients, of whom 2125 remained after quality control and stratification. Hence, only 14% of the patients in our final cohort had descriptive genetic information.

By adding biochemical, prescription, and genetic data we observed that the clusters were significantly different from each other on parameters other than comorbidities. By including the text narratives of the EHRs we were able to capture diagnoses that in another context would be considered as a primary diagnosis, for example epilepsy, schizophrenia and cerebral palsy. These diagnoses are not known comorbidities of diabetes but can influence the treatment and management of the diabetes patient. For instance, we observed that all patients in cluster 47 had schizophrenia, which could influence their compliance since the cluster was associated with glycemic dysregulation. We determined this when assessing the level of glycemic dysregulation and found that this cluster indeed showed a high number of patients with at least three parameters for glycemic dysregulation. However, a more in-depth analysis is required to clarify whether the glycemic dysregulation is due to the behavioral effects of schizophrenia, underlying genetic variants, adverse drug reactions due to polypharmacy, or other variables.

Despite our data from both assigned and text-mined diagnoses, misdiagnoses can occur, and we performed a manual inspection of randomly selected EHRs to establish the validity of the data. Furthermore, we observed some patients assigned with different diabetes types, for example first assigned with T1D and later with T2D, and vice versa. Inspecting the biochemical values of GAD65 autoantibodies and comparing them to the primary diagnosis type we found 182 T2D assigned individuals to have GAD65 levels above 10 IU/ml, possibly indicative of T1D or LADA; however, these individuals were not significantly enriched in any cluster. We also observed 621 individuals with GAD65 levels below 10 IU/ml, which is consistent with known late-term effects of T1D (results not shown). An in-depth temporal analysis of these patients with mixed diabetes types could be interesting and integrating biochemical as well as genetic variation data could elucidate which, if any, phenotype might be the most accurate.

In this study, we have used data from a unique cohort of 14,017 patients with diabetes, of which 12,866 had been diagnosed with either T1D or T2D. Integrating the assigned and text-mined ICD-10 and SDC-custom diagnoses, an MCL clustering was carried out which resulted in 172 unique clusters. Of these, 71 had at least 50 patients, which were subsequently divided into groups with shared symptoms. Investigating the complication enrichment and comorbidity patterns in the clusters and symptom groups we detected clusters described by specific disorders such as hypothyroidism, schizophrenia, and functional intestine disorder as well as unique comorbidity interaction patterns both with and without temporal significance. An interesting approach could be to extend the temporal analysis to investigate how disease progression within and between clusters and symptoms groups develops for multiple diagnoses. This could be done with a trajectory-based approach as done recently by Jensen et al. (2014).

Materials and methods

EHR data

Request a detailed protocol

All data originate from the Steno Diabetes Center Copenhagen (SDCC), a specialized diabetes hospital in the Capital Region of Denmark. In Denmark patients with type 1 diabetes (T1D) are followed in hospital outpatient clinics such as SDCC, and the T1D patients studied comprise 35% of all adult patients with T1D in the Capital Region of Denmark. Patients with type 2 diabetes (T2D) are referred from primary care for treatment optimization, typically for a period of six to twelve months. When treatment goals are reached, and they have no diabetic complications, they are referred back to general practice. Patients needing intensive control and treatment, because of micro- or macrovascular complications, are offered life-long follow-up at the SDCC. At any time, approximately 2000 patients with complicated T2D are followed at the SDCC. Generally, the patients registered in the SDCC electronic patient records are representative of Danish patients with T1D and the 10% most complicated patients with T2D (Jørgensen et al., 2016). Moreover, the patient followed at SDCC are comparable to patients followed in all Danish hospital diabetes outpatient clinics in terms of distribution of age and duration of diabetes. The data comprise all communications and contacts recorded at the hospital over a period of 19 years (1993–2012) for 14,017 patients. This includes, primary diagnoses, prescriptions and laboratory tests, 1.2M clinical narrative entries, 420 different types of laboratory tests with 4.15M laboratory measurements and a total number of 440,555 drug prescriptions. On average, each patient had 85 clinical narratives with an average length of 34 words (212 characters). In addition, genetic data from several research projects have been linked to the patients and added to the EHRs.

Text-mining dictionaries, tagging and corpus matches

Request a detailed protocol

An in-house developed framework for mining Danish text was used for the analysis (Roque et al., 2011; Eriksson et al., 2013). The algorithm tags words in the text narratives in a named entity recognition (NER) fashion based on supplied dictionaries. In this study, we used two main dictionaries: The International Classification of Disease version 10 (ICD-10) truncated to level 3 (e.g. E10: Type 1 Diabetes), and a complementary ‘exposome-oriented’ dictionary (SDC-custom). The latter holds terms related to diabetes specific subtypes (e.g. MODY and LADA), complications (e.g. the different severities of neuropathy, retinopathy and nephropathy), treatments and examinations (e.g. gastric bypass, renography, and beta blockers), lifestyle and lifestyle related disorders (e.g. obesity, exercise level, smoking), and compliance. The SDC-custom dictionary was developed in collaboration with physicians at the SDCC (see Supplementary file 6 for a translated and condensed version). The Danish ICD-10 version currently contains roughly 20,000 unique descriptions of clinical concepts, each with a unique ICD-10 code.

The NER used for dictionary matching, in addition, performs lemmatization and de-latinization of tagged words, accounts for language negations or subject negations (e.g. ‘the patient’s mother had retinopathy’), and performs fuzzy matching with a Hamming distance of 1 (e.g. ‘diabtes’ is transformed to its correct spelling ‘diabetes’). A thorough explanation of the algorithm is provided (Simon et al., 2019, manuscript in preparation). Other details, for example on ‘negation scope’, that is the position of negations relative to the negated term in Danish, have been published previously (Thomas et al., 2014).

Running the text-mining algorithm (Simon et al., 2019, manuscript in preparation) on the SDCC corpus with the two dictionaries (ICD-10 and SDC-custom) recognized 1,028,593 entities from the dictionaries in 12,504 patients (80.5% of the entire corpus). None of the remaining patients had any non-trivial match between the dictionaries and EHR narratives. The two dictionaries shared some general terms, for example T1D and T2D; these duplicate matches were removed and 941,087 unique code matches remained. Of these, 267,404 were fuzzy matches representing 4181 unique variants. The variants were manually validated, resulting in removal of 10,952 (4.1%) matches. After removal of negated sentences (n = 255,302) 594,600 code-to-text matches in 12,467 patients were left.

Patient phenotype vectors from assigned and text mined codes

Request a detailed protocol

The structured ICD-10 codes assigned to patients during their contact with SDCC were extracted from the EHRs, along with all ICD-10 codes captured by mining the text parts of the EHRs. The two ICD-10 lists were combined, but to prevent the primary, assigned diabetes types from dominating the patient stratification, diagnosis codes for diabetes without complications (E10 and E109, in total 3740 codes, and E11 and E119, in total 3624 codes) were removed. Approximately 8% of the assigned codes were removed in this way. The list of codes and their frequencies for each patient were transformed using the BM25 weighting scheme (Robertson and Walker, 1994), which scores a code c in patient P, accounting for the code frequency in all patients, frequency of the codes in the patient (document frequency), and number codes in the patient record (document length), see Equation 1.

(1) Score(p,c)=∑i−1nIDF(ci)⋅f(ci,p)⋅(k1+1)f(ci,p)+k1⋅(a−b+b⋅|p||pave|)

Here, IDF(c) is the inverse document frequency for the code c computed as

IDF(ci)=log⁡N−n(ci)+0.5n(ci)+0.5

With N being the total number of patients and n(c) the number of patients with a given code ci , and the term f(ci,p) is the frequency of code ci in patient p. The number of codes associated with each patient vector, P, is given by the length of the vector, |p|, and the average number of codes in the entire corpus is |pave|. Finally, b and k1 are free parameters that determine to what extent document length is considered (b) and how much the scoring equation resembles a normal TF-IDF (k1), respectively. The value of b was set to 0.75 and does not fully account for the document length (b=1) and k1 was set to 1.2 giving a low resemblance of TF-IDF (k1→∞).

Clustering patients from Cosine similarities

Request a detailed protocol

All patients were clustered using their pairwise cosine similarities calculated from the BM25 transformed code vectors. A cosine similarity ≥ 0.5 was set as a cut-off prior to clustering, to minimize the number of edges in the subsequent patient network. To increase the variance of the cosine similarities, these were scaled from the interval 0.5–1 to 10–100. We wanted to do a network based clustering and therefore used Markov Clustering (MCL) with the inflation parameter set to 1.2 and rest left as default (Van Dongen, 2000). Different inflation parameters were tested and evaluated based on the efficiency, mass fraction, and area fraction parameters.

Grouping clusters in symptom related groups

Request a detailed protocol

We organized the clusters into symptom groups based on the frequency of their symptom codes using ICD-10 chapter XVIII level 1, for example R50-69: General symptoms and signs. We used a Euclidean distance and applied a hierarchical clustering using Ward.D as the agglomeration method since we wanted to expose the hierarchical relationship amongst the clusters. The entire analysis was performed using R (version 3.2.1).

Enrichment analysis of diagnosis codes

Request a detailed protocol

The MCL clusters were tested for ICD-10 and SDC-custom codes found more often than expected, using a binominal test while correcting for sex and birth decade. The metadata such as average age, days at SDCC, and diabetes duration (from the date of diabetes diagnosis until the end of the study) were calculated, and further p-values for each cluster were obtained using a Wilcoxon test against the remaining clusters. In both analyses, p-values were adjusted using Benjamini-Hochberg correction for multiple testing, where a p-value≤0.05 was considered significant.

Comorbidity patterns for diagnosis pairs

Request a detailed protocol

We performed three independent analyzes without considering the clusters by applying Fischer’s exact tests to obtain p-values for all diagnosis pairs within the SDCC corpus: 1) p-values for observing the codes together, 2) p-values for observing diagnosis A prior to diagnosis B, and 3) p-values for observing diagnosis B prior to diagnosis A. P-values from the three different sets were adjusted using Bonferroni correction for multiple testing, and the pairs were subsequently ranked based on these values. To detect whether the pairs were observed more together than expected we applied a comorbidity score as described in Roque et al. (2011). For the temporal pairs, we also applied an ANOVA test to investigate whether any of these pairs were unique for a symptom group. All p-values were corrected for multiple testing, and an adjusted p-value≤0.05 was considered significant.

Robustness of the MCL generated clusters

Request a detailed protocol

To assess quantitatively the stability of the clusters generated, we constructed various diluted and shuffled realizations of the similarity network used as input to the MCL algorithm. We used a reference clustering similar to the clustering presented in Figure 2B (either by including the patients in the 71 clusters or all patients). The diluted versions were generated by randomly deleting edges with a probability of α, whereas the shuffled realizations were created by shuffling edges between nodes (patients) as described earlier (Karrer et al., 2008). The latter produces a network where the number of edges and vertices are unchanged. An α of zero leaves the reference network unchanged, while a value of 1 leads to a complete randomization of the similarity network. Each of these randomizations of the input were repeated five times for various values of α in the range 0–50% and used as input for the MCL algorithm. The resulting clustering’s were then compared to the reference clustering by means of the Variation of Information measure (VI) (Meilă, 2007) and plotted as function of increasing values of α (see Figure 2—figure supplement 5). The figure includes two horizontal lines corresponding to the value that the VI would take if we were to randomly assign 10% and 20% of the vertices to different random clusters, respectively. This analysis showed that the clustering is stable in relation to removing edges, which is evidence that the cosine metric-based cutoff used does not change the overall structure of the clustering. The shuffling is a more impactful randomization, however despite this, we can still shuffle around 10% of the edges and still retrieve 90% of the patients in the groups of the 71 reference clusters.

Quantitative assessment of glycemic dysregulation

Request a detailed protocol

Glycemic dysregulation was assessed for each patient by evaluating five different parameters. The first two parameters were obtained using the SDC-custom code for dysregulation (sdcL03) and the ICD-10 codes for hyperglycemia (R73 and E89). The remaining three were found by analyzing longitudinal measurements for glycated hemoglobin (HbA1c). Due to a large variation in both the number of measurements and their frequency, HbA1c values were pre-processed. We divided the HbA1c measurements for each patient into segments containing a minimum of five values, spanning a time interval of at least three months (equivalent to the functional lifetime of red blood cells). In total 10,112 patients had HbA1c measurements that fulfilled the criteria, and the subsequent analyses were performed on this sub-population.

We performed three analyses on the longitudinal pre-processed HbA1c data for each patient: 1) a Bayesian analysis of change point detection to find potential peaks of HbA1c values in a patient, 2) analysis of mixed effects models to estimate the HbA1c value at diabetes onset, and, 3) analysis of the frequency of values in different HbA1c bins (e.g. general level for diagnosing T1D or T2D, the critical interval for hyperglycemia etc.) to appoint an HbA1c severity score.

Laboratory test data

Request a detailed protocol

The laboratory tests were longitudinal data such as blood pressure measurements and biochemical analyses of blood and urine samples, and each test was assigned a unique identifier using the NPU-terminology, which is the recommended administration and communication measure of laboratory tests in Denmark (Petersen et al., 2012). In our data, several laboratory tests had an SDC identifier, being from local laboratory facilities at SDCC. Both test IDs, NPU and SDC, were analyzed separately, despite sometimes measuring the same biochemical variables.

In total, 420 different physiological tests were performed across 14,847 patients from the entire corpus. Measurements within and between tests were unbalanced with no general system in measurement interval, frequency, or number of patients who had a test taken. Due to this lack of systematic coverage, only tests that were performed on at least 75% of the entire corpus (10,788 patients) were analyzed (26 tests). However, the test for C-peptide (NPU18004) was also included as it was available for 74.9% of the cohort and is widely used to distinguish T1D and T2D. Measurements outside the biological reference interval for a given test, that is HbA1c measurements below 15 mmol/mol and above 184 mmol/mol, were removed, and for each patient the mean, median and standard deviation for each test with continuous values (20 of the 26 tests) were calculated. If the data was not normally distributed for a test we log-transformed it and normalized all values to mean = 0 and SD = 1. All analyses after assigning patients to clusters were performed on the 10,788 patients.

We applied a MANOVA to test if means among the three different patient groups (clusters, symptom groups or patients being dysregulated) were significantly different, and a Kolmogorov–Smirnov test was applied to investigate whether the distribution of the sample means in the patient groups were significantly higher or lower than means in the remaining groups. All p-values were adjusted using Bonferroni correction for multiple testing, and an adjusted p-value≤0.05 was considered significant.

Drug prescription data

Request a detailed protocol

Prescription data was available for 12,147 patients with a total number of 440,555 drug prescriptions. Drug compounds were identified by the ATC classification system, which is divided into groups at five different levels. In this study, we summarized the data using ATC-codes at level three and four: chemical and pharmacological and therapeutical, respectively.

From the initial set of prescriptions, we manually reviewed 104 drugs which did not have an ATC code in the EHR or were mapped to more than one ATC code. In addition to the manual review, pro.medicin (www.pro.medicin.dk, accessed October 2018) was used to map drug names to their corresponding ATC code. The SDCC prescription data and the WHO Collaborating Centre for Drug Statistics Methodology (www.whocc.no, accessed October 2018) were used for crosschecking. We performed Fisher’s exact test to investigate prescription enrichment (3rd level of the ATC classification) in clusters with at least 50 patients. The p-values were adjusted using Benjamini–Hochberg correction for multiple testing, and an adjusted p-value≤0.05 was considered significant.

Genomic characterization

Request a detailed protocol

A total of 2290 patients with T2D and 1028 patients with T1D from SDCC were genotyped separately using the HumanOmniExpress (24v1) array from Illumina as previously described (Charmet et al., 2018; Steinthorsdottir et al., 2014). Genotypes were called using GenomeStudio, and imputed separately using the Haplotype Reference Consortium (HRC) imputation panel (McCarthy et al., 2016). Prior to imputation, the two datasets were filtered to retain only high-quality samples/SNPs (sample call rate ≥98%, no mislabeled sex, no ethnic outliers, heterozygosity within 2 SD from the mean, SNP call rate ≥98%, no monomorphic SNPs, no Hardy–Weinberg disequilibrium outliers). After imputation, SNPs with minor allele frequency (MAF) <0.01, more than 20% missingness, R square less than 0.30, and duplicate SNPs were removed, and the two datasets were merged retaining only variants common to the two sets. After merging, relatedness between all individuals were calculated and close relatives were excluded. Of the 3318 patients, 2337 had EHR information and could be mapped to clusters. In total 2125 patients passed quality control and were taken forward for genomic characterization. Logistic regression was used to test for genetic differences (PLINK 1.90 beta, https://www.cog-genomics.org/1.9) between the different groups of interest (clusters and symptom groups) and linear regression was used to evaluate the SNPs impact on dysregulation. Cases were defined as all individuals in a given cluster/symptom group, and controls as all individuals not belonging to the respective cluster/symptom group. Glycemic dysregulation was defined as a score ranking from 0 (low) to 5 (high) based on five dysregulation parameters (see section on glycemic dysregulation). All analyses were adjusted for age and sex. The test statistics were adjusted for inflation (population stratification) using the three first principal components estimated using the --pca function in PLINK. Genetic associations were defined based on data derived from the EBI GWAS catalog version 1.0.1 (http://www.ebi.ac.uk/gwas/) unless otherwise stated. A p-value less than 5*10–8 was considered genome-wide significant.

Data availability

All data generated or analysed during this study are included in the manuscript and supporting files except for the raw person sensitive electronic health record data due to confidentiality requirements.

References

    1. Chen Y
    2. Li L
    3. Xu R
    (2015)
    Disease comorbidity network guides the detection of molecular evidence for the link between colorectal Cancer and obesity
    AMIA Joint Summits on Translational Science Proceedings. AMIA Joint Summits on Translational Science 2015:201–206.
    1. Hunt DL
    (2011)
    Diabetes: foot ulcers and amputations
    BMJ Clinical Evidence 2011:0602.
    1. Jones GT
    2. Tromp G
    3. Kuivaniemi H
    4. Gretarsdottir S
    5. Baas AF
    6. Giusti B
    7. Strauss E
    8. Van't Hof FN
    9. Webb TR
    10. Erdman R
    11. Ritchie MD
    12. Elmore JR
    13. Verma A
    14. Pendergrass S
    15. Kullo IJ
    16. Ye Z
    17. Peissig PL
    18. Gottesman O
    19. Verma SS
    20. Malinowski J
    21. Rasmussen-Torvik LJ
    22. Borthwick KM
    23. Smelser DT
    24. Crosslin DR
    25. de Andrade M
    26. Ryer EJ
    27. McCarty CA
    28. Böttinger EP
    29. Pacheco JA
    30. Crawford DC
    31. Carrell DS
    32. Gerhard GS
    33. Franklin DP
    34. Carey DJ
    35. Phillips VL
    36. Williams MJ
    37. Wei W
    38. Blair R
    39. Hill AA
    40. Vasudevan TM
    41. Lewis DR
    42. Thomson IA
    43. Krysa J
    44. Hill GB
    45. Roake J
    46. Merriman TR
    47. Oszkinis G
    48. Galora S
    49. Saracini C
    50. Abbate R
    51. Pulli R
    52. Pratesi C
    53. Saratzis A
    54. Verissimo AR
    55. Bumpstead S
    56. Badger SA
    57. Clough RE
    58. Cockerill G
    59. Hafez H
    60. Scott DJ
    61. Futers TS
    62. Romaine SP
    63. Bridge K
    64. Griffin KJ
    65. Bailey MA
    66. Smith A
    67. Thompson MM
    68. van Bockxmeer FM
    69. Matthiasson SE
    70. Thorleifsson G
    71. Thorsteinsdottir U
    72. Blankensteijn JD
    73. Teijink JA
    74. Wijmenga C
    75. de Graaf J
    76. Kiemeney LA
    77. Lindholt JS
    78. Hughes A
    79. Bradley DT
    80. Stirrups K
    81. Golledge J
    82. Norman PE
    83. Powell JT
    84. Humphries SE
    85. Hamby SE
    86. Goodall AH
    87. Nelson CP
    88. Sakalihasan N
    89. Courtois A
    90. Ferrell RE
    91. Eriksson P
    92. Folkersen L
    93. Franco-Cereceda A
    94. Eicher JD
    95. Johnson AD
    96. Betsholtz C
    97. Ruusalepp A
    98. Franzén O
    99. Schadt EE
    100. Björkegren JL
    101. Lipovich L
    102. Drolet AM
    103. Verhoeven EL
    104. Zeebregts CJ
    105. Geelkerken RH
    106. van Sambeek MR
    107. van Sterkenburg SM
    108. de Vries JP
    109. Stefansson K
    110. Thompson JR
    111. de Bakker PI
    112. Deloukas P
    113. Sayers RD
    114. Harrison SC
    115. van Rij AM
    116. Samani NJ
    117. Bown MJ
    (2017) Meta-Analysis of Genome-Wide association studies for abdominal aortic aneurysm identifies four new Disease-Specific risk loci
    Circulation Research 120:341–353.
    https://doi.org/10.1161/CIRCRESAHA.116.308765
  48. Conference
    1. Robertson SE
    2. Walker S
    (1994)
    Some simple effective approximations to the 2–Poisson Model for Probabilistic Weighted Retrieval
    SIGIR ’94 Proceedings of the 17th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval. pp. 232–241.
    1. Teslovich TM
    2. Musunuru K
    3. Smith AV
    4. Edmondson AC
    5. Stylianou IM
    6. Koseki M
    7. Pirruccello JP
    8. Ripatti S
    9. Chasman DI
    10. Willer CJ
    11. Johansen CT
    12. Fouchier SW
    13. Isaacs A
    14. Peloso GM
    15. Barbalic M
    16. Ricketts SL
    17. Bis JC
    18. Aulchenko YS
    19. Thorleifsson G
    20. Feitosa MF
    21. Chambers J
    22. Orho-Melander M
    23. Melander O
    24. Johnson T
    25. Li X
    26. Guo X
    27. Li M
    28. Shin Cho Y
    29. Jin Go M
    30. Jin Kim Y
    31. Lee JY
    32. Park T
    33. Kim K
    34. Sim X
    35. Twee-Hee Ong R
    36. Croteau-Chonka DC
    37. Lange LA
    38. Smith JD
    39. Song K
    40. Hua Zhao J
    41. Yuan X
    42. Luan J
    43. Lamina C
    44. Ziegler A
    45. Zhang W
    46. Zee RY
    47. Wright AF
    48. Witteman JC
    49. Wilson JF
    50. Willemsen G
    51. Wichmann HE
    52. Whitfield JB
    53. Waterworth DM
    54. Wareham NJ
    55. Waeber G
    56. Vollenweider P
    57. Voight BF
    58. Vitart V
    59. Uitterlinden AG
    60. Uda M
    61. Tuomilehto J
    62. Thompson JR
    63. Tanaka T
    64. Surakka I
    65. Stringham HM
    66. Spector TD
    67. Soranzo N
    68. Smit JH
    69. Sinisalo J
    70. Silander K
    71. Sijbrands EJ
    72. Scuteri A
    73. Scott J
    74. Schlessinger D
    75. Sanna S
    76. Salomaa V
    77. Saharinen J
    78. Sabatti C
    79. Ruokonen A
    80. Rudan I
    81. Rose LM
    82. Roberts R
    83. Rieder M
    84. Psaty BM
    85. Pramstaller PP
    86. Pichler I
    87. Perola M
    88. Penninx BW
    89. Pedersen NL
    90. Pattaro C
    91. Parker AN
    92. Pare G
    93. Oostra BA
    94. O'Donnell CJ
    95. Nieminen MS
    96. Nickerson DA
    97. Montgomery GW
    98. Meitinger T
    99. McPherson R
    100. McCarthy MI
    101. McArdle W
    102. Masson D
    103. Martin NG
    104. Marroni F
    105. Mangino M
    106. Magnusson PK
    107. Lucas G
    108. Luben R
    109. Loos RJ
    110. Lokki ML
    111. Lettre G
    112. Langenberg C
    113. Launer LJ
    114. Lakatta EG
    115. Laaksonen R
    116. Kyvik KO
    117. Kronenberg F
    118. König IR
    119. Khaw KT
    120. Kaprio J
    121. Kaplan LM
    122. Johansson A
    123. Jarvelin MR
    124. Janssens AC
    125. Ingelsson E
    126. Igl W
    127. Kees Hovingh G
    128. Hottenga JJ
    129. Hofman A
    130. Hicks AA
    131. Hengstenberg C
    132. Heid IM
    133. Hayward C
    134. Havulinna AS
    135. Hastie ND
    136. Harris TB
    137. Haritunians T
    138. Hall AS
    139. Gyllensten U
    140. Guiducci C
    141. Groop LC
    142. Gonzalez E
    143. Gieger C
    144. Freimer NB
    145. Ferrucci L
    146. Erdmann J
    147. Elliott P
    148. Ejebe KG
    149. Döring A
    150. Dominiczak AF
    151. Demissie S
    152. Deloukas P
    153. de Geus EJ
    154. de Faire U
    155. Crawford G
    156. Collins FS
    157. Chen YD
    158. Caulfield MJ
    159. Campbell H
    160. Burtt NP
    161. Bonnycastle LL
    162. Boomsma DI
    163. Boekholdt SM
    164. Bergman RN
    165. Barroso I
    166. Bandinelli S
    167. Ballantyne CM
    168. Assimes TL
    169. Quertermous T
    170. Altshuler D
    171. Seielstad M
    172. Wong TY
    173. Tai ES
    174. Feranil AB
    175. Kuzawa CW
    176. Adair LS
    177. Taylor HA
    178. Borecki IB
    179. Gabriel SB
    180. Wilson JG
    181. Holm H
    182. Thorsteinsdottir U
    183. Gudnason V
    184. Krauss RM
    185. Mohlke KL
    186. Ordovas JM
    187. Munroe PB
    188. Kooner JS
    189. Tall AR
    190. Hegele RA
    191. Kastelein JJ
    192. Schadt EE
    193. Rotter JI
    194. Boerwinkle E
    195. Strachan DP
    196. Mooser V
    197. Stefansson K
    198. Reilly MP
    199. Samani NJ
    200. Schunkert H
    201. Cupples LA
    202. Sandhu MS
    203. Ridker PM
    204. Rader DJ
    205. van Duijn CM
    206. Peltonen L
    207. Abecasis GR
    208. Boehnke M
    209. Kathiresan S
    (2010) Biological, clinical and population relevance of 95 loci for blood lipids
    Nature 466:707–713.
    https://doi.org/10.1038/nature09270
  58. Conference
    1. Thomas CE
    2. Bjødstrup Jensen P
    3. Werge T
    4. Brunak S
    (2014)
    Negation scope and spelling variation for text-mining of Danish electronic patient records
    Proceedings of the 5th International Workshop on Health Text Mining and Information Analysis. pp. 64–88.
  62. Thesis
    1. Van Dongen S
    (2000)
    Graph Clustering by Flow Simulation
    University of Utrecht.

Article and author information

Author details

  1. Isa Kristina Kirk

    Novo Nordisk Foundation Center for Protein Research, University of Copenhagen, Copenhagen, Denmark
    Contribution
    Conceptualization, Resources, Data curation, Software, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing—original draft, Project administration, Writing—review and editing
    Contributed equally with
    Christian Simon
    Competing interests
    No competing interests declared

  2. Christian Simon

    Novo Nordisk Foundation Center for Protein Research, University of Copenhagen, Copenhagen, Denmark
    Contribution
    Conceptualization, Resources, Data curation, Software, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing—original draft, Project administration, Writing—review and editing
    Contributed equally with
    Isa Kristina Kirk
    Competing interests
    No competing interests declared

  3. Karina Banasik

    Novo Nordisk Foundation Center for Protein Research, University of Copenhagen, Copenhagen, Denmark
    Contribution
    Formal analysis, Methodology, Writing—original draft, Project administration, Writing—review and editing
    Competing interests
    No competing interests declared

  4. Peter Christoffer Holm

    Novo Nordisk Foundation Center for Protein Research, University of Copenhagen, Copenhagen, Denmark
    Contribution
    Formal analysis, Methodology, Writing—review and editing
    Competing interests
    No competing interests declared

  5. Amalie Dahl Haue

    Novo Nordisk Foundation Center for Protein Research, University of Copenhagen, Copenhagen, Denmark
    Contribution
    Formal analysis, Methodology, Writing—review and editing
    Competing interests
    No competing interests declared

  6. Peter Bjødstrup Jensen

    1. Novo Nordisk Foundation Center for Protein Research, University of Copenhagen, Copenhagen, Denmark
    2. Odense Patient Data Explorative Network (OPEN), Odense University Hospital, Odense, Denmark
    Contribution
    Formal analysis, Supervision, Methodology, Writing—review and editing
    Competing interests
    No competing interests declared

  7. Lars Juhl Jensen

    Novo Nordisk Foundation Center for Protein Research, University of Copenhagen, Copenhagen, Denmark
    Contribution
    Conceptualization, Data curation, Formal analysis, Supervision, Methodology, Writing—original draft, Writing—review and editing
    Competing interests
    No competing interests declared

  8. Cristina Leal Rodríguez

    Novo Nordisk Foundation Center for Protein Research, University of Copenhagen, Copenhagen, Denmark
    Contribution
    Methodology, Writing—review and editing
    Competing interests
    No competing interests declared

  9. Mette Krogh Pedersen

    Novo Nordisk Foundation Center for Protein Research, University of Copenhagen, Copenhagen, Denmark
    Contribution
    Data curation, Methodology
    Competing interests
    No competing interests declared

  10. Robert Eriksson

    Novo Nordisk Foundation Center for Protein Research, University of Copenhagen, Copenhagen, Denmark
    Contribution
    Data curation, Methodology
    Competing interests
    No competing interests declared

  11. Henrik Ullits Andersen

    Steno Diabetes Center Copenhagen, Gentofte, Denmark
    Contribution
    Conceptualization, Resources, Data curation, Validation, Writing—review and editing
    Competing interests
    No competing interests declared

  12. Thomas Almdal

    1. Steno Diabetes Center Copenhagen, Gentofte, Denmark
    2. Department of Endocrinology, Rigshospitalet, Copenhagen, Denmark
    Contribution
    Conceptualization, Resources, Data curation, Writing—review and editing
    Competing interests
    No competing interests declared

  13. Jette Bork-Jensen

    Novo Nordisk Foundation Center for Basic Metabolic Research, University of Copenhagen, Copenhagen, Denmark
    Contribution
    Data curation, Writing—review and editing
    Competing interests
    No competing interests declared

  14. Niels Grarup

    Novo Nordisk Foundation Center for Basic Metabolic Research, University of Copenhagen, Copenhagen, Denmark
    Contribution
    Data curation, Writing—review and editing
    Competing interests
    No competing interests declared

  15. Knut Borch-Johnsen

    Holbæk Hospital, Holbæk, Denmark
    Contribution
    Conceptualization, Resources, Supervision, Project administration, Writing—review and editing
    Competing interests
    No competing interests declared

  16. Oluf Pedersen

    1. Steno Diabetes Center Copenhagen, Gentofte, Denmark
    2. Novo Nordisk Foundation Center for Basic Metabolic Research, University of Copenhagen, Copenhagen, Denmark
    Contribution
    Conceptualization, Resources, Data curation, Writing—review and editing
    Competing interests
    No competing interests declared

  17. Flemming Pociot

    1. Steno Diabetes Center Copenhagen, Gentofte, Denmark
    2. Department of Clinical Medicine, Herlev-Gentofte Hospital, Herlev, Denmark
    Contribution
    Resources, Data curation, Methodology, Writing—review and editing
    Competing interests
    No competing interests declared

  18. Torben Hansen

    1. Steno Diabetes Center Copenhagen, Gentofte, Denmark
    2. Novo Nordisk Foundation Center for Basic Metabolic Research, University of Copenhagen, Copenhagen, Denmark
    Contribution
    Conceptualization, Resources, Data curation, Supervision, Writing—review and editing
    Competing interests
    No competing interests declared

  19. Regine Bergholdt

    Steno Diabetes Center Copenhagen, Gentofte, Denmark
    Contribution
    Conceptualization, Resources, Data curation, Methodology, Project administration, Writing—review and editing
    Competing interests
    No competing interests declared

  20. Peter Rossing

    1. Steno Diabetes Center Copenhagen, Gentofte, Denmark
    2. Department of Clinical Medicine, University of Copenhagen, Copenhagen, Denmark
    Contribution
    Conceptualization, Resources, Data curation, Supervision, Methodology, Writing—original draft, Project administration, Writing—review and editing
    For correspondence
    peter.rossing@regionh.dk
    Competing interests
    No competing interests declared

  21. Søren Brunak

    1. Novo Nordisk Foundation Center for Protein Research, University of Copenhagen, Copenhagen, Denmark
    2. Center for Biological Sequence Analysis, Department of Bio and Health Informatics, Technical University of Denmark, Lyngby, Denmark
    Contribution
    Conceptualization, Resources, Data curation, Funding acquisition, Methodology, Writing—original draft, Project administration, Writing—review and editing
    For correspondence
    soren.brunak@cpr.ku.dk
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-0316-5866

Funding

Danish Council for Strategic Research (0603-00321B)

  • Søren Brunak

Innovation Fund Denmark (5153-00002B)

  • Søren Brunak

Novo Nordisk Foundation (NNF14CC0001)

  • Søren Brunak

Novo Nordisk Foundation (NNF17OC0027594)

  • Søren Brunak

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Copyright

© 2019, Kirk et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 1,467
    views
  • 151
    downloads
  • 10
    citations

Views, downloads and citations are aggregated across all versions of this paper published by eLife.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

  1. Isa Kristina Kirk
  2. Christian Simon
  3. Karina Banasik
  4. Peter Christoffer Holm
  5. Amalie Dahl Haue
  6. Peter Bjødstrup Jensen
  7. Lars Juhl Jensen
  8. Cristina Leal Rodríguez
  9. Mette Krogh Pedersen
  10. Robert Eriksson
  11. Henrik Ullits Andersen
  12. Thomas Almdal
  13. Jette Bork-Jensen
  14. Niels Grarup
  15. Knut Borch-Johnsen
  16. Oluf Pedersen
  17. Flemming Pociot
  18. Torben Hansen
  19. Regine Bergholdt
  20. Peter Rossing
  21. Søren Brunak
(2019)
Linking glycemic dysregulation in diabetes to symptoms, comorbidities, and genetics through EHR data mining
eLife 8:e44941.
https://doi.org/10.7554/eLife.44941

Share this article

https://doi.org/10.7554/eLife.44941

Categories and tags

Research organism

Further reading

    1. Cancer Biology
    2. Computational and Systems Biology

    Luminal epithelial cells integrate variable responses to aging into stereotypical changes that underlie breast cancer susceptibility

    Rosalyn W Sayaman, Masaru Miyano ... Mark LaBarge
    Research Article

    Effects from aging in single cells are heterogenous, whereas at the organ- and tissue-levels aging phenotypes tend to appear as stereotypical changes. The mammary epithelium is a bilayer of two major phenotypically and functionally distinct cell lineages: luminal epithelial and myoepithelial cells. Mammary luminal epithelia exhibit substantial stereotypical changes with age that merit attention because these cells are the putative cells-of-origin for breast cancers. We hypothesize that effects from aging that impinge upon maintenance of lineage fidelity increase susceptibility to cancer initiation. We generated and analyzed transcriptomes from primary luminal epithelial and myoepithelial cells from younger <30 (y)ears old and older >55y women. In addition to age-dependent directional changes in gene expression, we observed increased transcriptional variance with age that contributed to genome-wide loss of lineage fidelity. Age-dependent variant responses were common to both lineages, whereas directional changes were almost exclusively detected in luminal epithelia and involved altered regulation of chromatin and genome organizers such as SATB1. Epithelial expression of gap junction protein GJB6 increased with age, and modulation of GJB6 expression in heterochronous co-cultures revealed that it provided a communication conduit from myoepithelial cells that drove directional change in luminal cells. Age-dependent luminal transcriptomes comprised a prominent signal that could be detected in bulk tissue during aging and transition into cancers. A machine learning classifier based on luminal-specific aging distinguished normal from cancer tissue and was highly predictive of breast cancer subtype. We speculate that luminal epithelia are the ultimate site of integration of the variant responses to aging in their surrounding tissue, and that their emergent phenotype both endows cells with the ability to become cancer-cells-of-origin and represents a biosensor that presages cancer susceptibility.

    1. Computational and Systems Biology
    2. Microbiology and Infectious Disease

    An antimicrobial drug recommender system using MALDI-TOF MS and dual-branch neural networks

    Gaetan De Waele, Gerben Menschaert, Willem Waegeman
    Research Article

    Timely and effective use of antimicrobial drugs can improve patient outcomes, as well as help safeguard against resistance development. Matrix-assisted laser desorption/ionization time-of-flight mass spectrometry (MALDI-TOF MS) is currently routinely used in clinical diagnostics for rapid species identification. Mining additional data from said spectra in the form of antimicrobial resistance (AMR) profiles is, therefore, highly promising. Such AMR profiles could serve as a drop-in solution for drastically improving treatment efficiency, effectiveness, and costs. This study endeavors to develop the first machine learning models capable of predicting AMR profiles for the whole repertoire of species and drugs encountered in clinical microbiology. The resulting models can be interpreted as drug recommender systems for infectious diseases. We find that our dual-branch method delivers considerably higher performance compared to previous approaches. In addition, experiments show that the models can be efficiently fine-tuned to data from other clinical laboratories. MALDI-TOF-based AMR recommender systems can, hence, greatly extend the value of MALDI-TOF MS for clinical diagnostics. All code supporting this study is distributed on PyPI and is packaged at https://github.com/gdewael/maldi-nn.

    1. Computational and Systems Biology
    2. Genetics and Genomics

    Functional characteristics and computational model of abundant hyperactive loci in the human genome

    Sanjarbek Hudaiberdiev, Ivan Ovcharenko
    Research Article

    Enhancers and promoters are classically considered to be bound by a small set of transcription factors (TFs) in a sequence-specific manner. This assumption has come under increasing skepticism as the datasets of ChIP-seq assays of TFs have expanded. In particular, high-occupancy target (HOT) loci attract hundreds of TFs with often no detectable correlation between ChIP-seq peaks and DNA-binding motif presence. Here, we used a set of 1003 TF ChIP-seq datasets (HepG2, K562, H1) to analyze the patterns of ChIP-seq peak co-occurrence in combination with functional genomics datasets. We identified 43,891 HOT loci forming at the promoter (53%) and enhancer (47%) regions. HOT promoters regulate housekeeping genes, whereas HOT enhancers are involved in tissue-specific process regulation. HOT loci form the foundation of human super-enhancers and evolve under strong negative selection, with some of these loci being located in ultraconserved regions. Sequence-based classification analysis of HOT loci suggested that their formation is driven by the sequence features, and the density of mapped ChIP-seq peaks across TF-bound loci correlates with sequence features and the expression level of flanking genes. Based on the affinities to bind to promoters and enhancers we detected five distinct clusters of TFs that form the core of the HOT loci. We report an abundance of HOT loci in the human genome and a commitment of 51% of all TF ChIP-seq binding events to HOT locus formation thus challenging the classical model of enhancer activity and propose a model of HOT locus formation based on the existence of large transcriptional condensates.