Data-driven identification of temporal glucose patterns in a large cohort of nondiabetic patients with COVID-19 using time-series clustering.

OBJECTIVE: Hyperglycemia has emerged as an important clinical manifestation of coronavirus disease 2019 (COVID-19) in diabetic and nondiabetic patients. Whether these glycemic changes are specific to a subgroup of patients and persist following COVID-19 resolution remains to be elucidated. This work aimed to characterize longitudinal random blood glucose in a large cohort of nondiabetic patients diagnosed with COVID-19.
MATERIALS AND METHODS: De-identified electronic medical records of 7502 patients diagnosed with COVID-19 without prior diagnosis of diabetes between January 1, 2020, and November 18, 2020, were accessed through the TriNetX Research Network. Glucose measurements, diagnostic codes, medication codes, laboratory values, vital signs, and demographics were extracted before, during, and after COVID-19 diagnosis. Unsupervised time-series clustering algorithms were trained to identify distinct clusters of glucose trajectories. Cluster associations were tested for demographic variables, COVID-19 severity, glucose-altering medications, glucose values, and new-onset diabetes diagnoses.
RESULTS: Time-series clustering identified a low-complexity model with 3 clusters and a high-complexity model with 19 clusters as the best-performing models. In both models, cluster membership differed significantly by death status, COVID-19 severity, and glucose levels. Clusters membership in the 19 cluster model also differed significantly by age, sex, and new-onset diabetes mellitus. DISCUSSION AND
CONCLUSION: This work identified distinct longitudinal blood glucose changes associated with subclinical glucose dysfunction in the low-complexity model and increased new-onset diabetes incidence in the high-complexity model. Together, these findings highlight the utility of data-driven techniques to elucidate longitudinal glycemic dysfunction in patients with COVID-19 and provide clinical evidence for further evaluation of the role of COVID-19 in diabetes pathogenesis.

INTRODUCTION

Coronavirus disease 2019 (COVID-19) is caused by infection with severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) and has resulted in over 3 million deaths worldwide as of April 2021. Hyperglycemia, defined as elevated blood glucose, has emerged as an important manifestation of COVID-19 in diabetic and nondiabetic patients. Studies have linked hyperglycemia to worse clinical outcomes,, COVID-19 severity, exacerbation of diabetes symptoms, and mortality,, but few have explored whether these glycemic changes persist following resolution of COVID-19.

Recent case reports of new-onset diabetes,, have increased speculation of long-term glycemic dysfunction following COVID-19 infection. Coronaviruses have previously been associated with dysregulation of glucose metabolism and transient hyperglycemia. SARS-CoV-2 is thought to mediate impaired beta-cell function through angiotensin-converting enzyme 2 receptor-induced inflammatory responses in the pancreas., SARS-CoV-2-associated inflammation may also impair insulin signaling in adipose tissue, resulting in hyperglycemia and insulin resistance. Recent findings suggest a pathogenic role of COVID-19 in inducing diabetes through beta-cell transdifferentiation and direct beta-cell apoptosis. Together, these findings indicate potential mechanisms for persistent COVID-19 induced glucose dysfunction.

Evidence evaluating longitudinal effects of COVID-19 on blood glucose has been difficult to ascertain. Considerable heterogeneity observed in patients who experienced hyperglycemia during COVID-19 suggests that there may be subgroups of patients with similar glucose changes. Reported incidences of new-onset diabetes mellitus also varied, as some studies found an increased prevalence during peak periods of the pandemic, others reported no change or a decrease. The CoviDIAB Project was established to define the phenotype of COVID-19-related new-onset diabetes, but does not consider transient or subclinical instances of glucose dysfunction. Previous studies with small sample sizes and short time frames of analysis highlight the need for big data repositories in evaluating long-term effects of COVID-19 and identifying subphenotypes of glucose dysregulation.

This work aims to characterize longitudinal glucose changes in a large cohort of patients with COVID-19. De-identified electronic medical records were utilized to train unsupervised machine learning models to identify glucose trajectories. The goal of this work was to provide a foundational framework for data-driven methods in identifying clinically useful stratifications of blood glucose.

METHODS

In this section, we briefly describe the methods used in this work. For a detailed and reproducible account, see Supplementary Methods.

Study cohort

TriNetX, a health research network whose functionalities are described elsewhere, (Supplementary Methods), was queried for COVID-19 patients confirmed by ICD-10 codes or laboratory testing indicating the presence of SARS-CoV-2 between January 20, 2020, and November 18, 2020 (Supplementary Table S1). De-identified data were obtained for 256,566 patients with confirmed COVID-19. Patients with at least 1 glucose measurement defined by laboratory codes (Supplementary Table S1) in each of the following timepoints were included: before (2 years to 1 week before COVID-19 diagnosis), during (1 week before to 2 weeks after COVID-19 diagnosis), and after (2 weeks to 1 year after COVID-19 diagnosis). Patients with a history of type 1 or type 2 diabetes defined by ICD-9 or ICD-10 codes prior to COVID-19 diagnosis were excluded (Supplementary Table S1). Using these inclusion and exclusion criteria, a cohort of 7502 patients was assembled (Supplementary Figure S1). The study was deemed exempt by the University of Utah Institutional Review Board.

Preprocessing

Random blood glucose measurements were extracted using LOINC value sets for glucose, and 3 trajectory datasets were created: segment, 3-month, and 6-month (Supplementary Methods). The segment trajectory dataset was created by averaging glucose at each timepoint (before, during, and after). The 3-month trajectory dataset was created by averaging glucose over 3 months’ intervals at each timepoint. The 6-month trajectory dataset was created by averaging glucose over 6 months’ intervals at each timepoint. All trajectory datasets were scaled using a mean-variance scaler (tslearn 0.4.1 in Python).

Clustering

Unsupervised k-means clustering was performed to identify distinct glucose trajectories (Supplementary Methods). In this work, 2 k-means unsupervised algorithms were trained: time-series k-means (ts k-means) and kernel k-means, (kernel) (tslearn 0.4.1 in Python). Clustering models were created for each unique combination of trajectory datasets (segment, 3-month, and 6-month) and clustering algorithm (ts k-means and kernel) (Supplementary Figure S2). Clustering was performed for k of 2–20 and performance was evaluated using the Silhouette Score (SS) (scikit-learn 0.24.1 in Python,) (Supplementary Methods). The model with the best SS score as confirmed by elbow method heuristics was selected for further analysis. Trajectories assigned to each cluster and cluster centers were plotted.

Phenotype analysis

Data extraction

Medications, diagnoses, vital signs, and demographics facts were downloaded for all 7502 patients in the cohort (Supplementary Figure S3). Standardized terminologies and terminology harmonization methods are described in the Supplementary Methods.

luster interpretation

Clusters with similar centers were grouped and features were extracted for each group (Supplementary Methods). Categorical features included the proportion of patients with age older than 65, female sex, death, moderate disease severity, severe disease severity, critical disease severity, hyperglycemic before, during, and after COVID-19, antihyperglycemic agents and insulin, glucose, steroids, and new-onset type 1 or type 2 diabetes (Supplementary Table S2). Continuous features included age, age at death, glucose before, during, and after COVID-19 diagnosis.

Statistical analysis

Continuous features were assessed for normality using D’Agostino’s test, (scipy 1.6.0 in Python) and independence using Kruskal–Wallis H-test (Pingouin 0.3.9 in Python). Categorical features were assessed for independence using Chi-square tests with Yates continuity correction (Pingouin 0.3.9 in Python). Categorical variables were also evaluated for Cramer’s V effect size (Pingouin 0.3.9 in Python). P-values for continuous and categorical variables were considered significant at the .05 level. Statistically significant categorical features were further analyzed by scaling the proportions for each feature and visualized using a radar diagram (plotly v4.14.3 in Python).

RESULTS

Cohort characteristics

For the 7502 patients with COVID-19, the mean (standard deviation [SD]) age was 56.25 (19.25) years, 56% (n = 4197) were female, 7% (n = 514) died, and the mean (SD) death age was 66.72 (17.50) years. From the cohort, 8% (n = 578) of patients were classified as having critical, 3% (n = 244) as severe, and 4% (n = 277) as moderate COVID-19 severity; 7% (n = 494) patients were prescribed antihyperglycemic agents and insulin, 12% (n = 915) were prescribed glucose, and 14% (n = 1080) were prescribed steroids.

The mean (SD) glucose before COVID-19 diagnosis was 6.23 (2.01) mmol/L and 21% (1592) were hyperglycemic. The mean (SD) glucose during COVID-19 diagnosis was 6.06 (1.71) mmol/L and 14% (n = 1084) were hyperglycemic. The mean (SD) glucose after COVID-19 diagnosis was 6.42 (1.97) mmol/L and 24% (n = 1824) were hyperglycemic; 1.4% (n = 103) of patients diagnosed developed new-onset diabetes during or after COVID-19.

Clustering results

A total of 114 models with unique trajectories (segment, 3-month, or 6-month) and algorithms (ts k-means or kernel) were trained on clusters ranging from 2 to 20. The comparison of the best-performing clusters from each model (Supplementary Table S4) clearly shows the superiority of the segment approach. The model with the best SS corresponds to the segment trajectory dataset, ts k-means, and 19 clusters (SS = 0.53). However, elbow method evaluation of the segment trajectory dataset and ts k-means models demonstrated a positive inflection point at 3 clusters (SS = 0.52) (Supplementary Figure S4). Therefore, both models were selected for further analysis as exemplars of low-complexity with k = 3 and high-complexity clustering with k = 19.

Three clusters model

Cluster visualization

Cluster trajectories and centers from the 3 cluster model, segment trajectory dataset, and ts k-means model are depicted in Figure 1. Three patterns were visually identified according to the cluster center trajectory. Cluster 1 was labeled as “Risers” because its center demonstrated an increasing pattern in scaled glucose during and after COVID-19 diagnosis (Figure 1A, top). Cluster 2 was labeled as “Decliners” because its center demonstrated a decreasing pattern in scaled glucose during and after COVID-19 diagnosis (Figure 1B, top). Cluster 3 was labeled as “Peakers” because its center demonstrated a peak in scaled glucose only during COVID-19 diagnosis (Figure 1C, top).

Figure 1.

Glucose trajectories and radar plots of phenotypic features of 3 clusters from the segment and k-means model. (A) Glucose trajectory of the ‘Risers’ cluster with black lines representing individual trajectories and the red line representing cluster centers (top) and radar plot of the scaled statistically significant features (bottom). (B) Glucose trajectory of the ‘Decliners’ cluster with black lines representing individual trajectories and the red line representing cluster centers (top) and radar plot of the scaled statistically significant features (bottom). (C) Glucose trajectory of the ‘Peakers’ cluster with black lines representing individual trajectories and the red line representing cluster centers (top) and radar plot of the scaled statistically significant features (bottom).

Phenotype results

Summary statistics for all features were calculated for each cluster (Table 1).

Table 1.

Statistical summary of phenotypic features of 3 clusters from the segment and k-means model

Feature		Cluster 1	Cluster 2	Cluster 3	P-value
		Risers (n = 2976)	Decliners (n = 2341)	Peakers (n = 2185)
Demographics
Age	Median (IQR) (years)	59 (27)	58 (29)	58 (30)	P = .082
	% >65 years (n)	36% (1085)	34% (802)	34% (738)	P = .091
Sex	% Female (n)	56% (1657)	56% (1312)	56% (1228)	P = .926
Death	Median (IQR) (years)	70 (22)	71 (21)	70 (24.5)	P = .798
	% Death (n)	6% (193)	10% (233)	4% (88)	P < .001
					V = .092
COVID-19 severity
Moderate	% (n)	4% (130)	4% (82)	3% (65)	P = .027
					V = .031
Severe	% (n)	4% (111)	3% (65)	3% (68)	P = .137
Critical	% (n)	9% (276)	8% (183)	5% (119)	P < .001
					V = .059
Glucose-altering medications
Anti-hyperglycemic agents and insulin	% (n)	7% (208)	7% (153)	6% (133)	P = .432
Glucose	% (n)	12% (361)	13% (293)	12% (261)	P = .833
Steroids	% (n)	15% (449)	15% (342)	13% (289)	P = .160
Glucose levels
Before	Median (IQR) (mmol/L)	5.55 (1.28)	6.60 (1.83)	5.28 (1.05)	P < .001
	% Hyperglycemic (n)	15% (445)	40% (925)	9% (201)	P < .001
					V = .313
During	Median (IQR) (mmol/L)	5.55 (1.05)	5.55 (1.00)	6.16 (1.39)	P < .001
	% Hyperglycemic (n)	10% (290)	10% (223)	26% (561)	P < .001
					V = .208
After	Median (IQR) (mmol/L)	6.66 (1.94)	5.72 (1.33)	5.44 (1.11)	P < .001
	% Hyperglycemic (n)	40% (1191)	17% (395)	10% (219)	P < .001
					V = .309
New-onset diabetes diagnoses
Diabetes	% (n)	1% (40)	1% (34)	1% (29)	P = .923

Note: Statistical summary of all features is presented for each cluster. Differences in continuous variables were tested using Kruskal–Wallis H-test. Differences between clusters in categorical variables were tested using the Chi-squared test for independence with Yates Continuity Correction and Cramer’s V for effect size was calculated for statistically significant features. P-values were evaluated at the .05 significance level, and the bold text indicates statistical significance.

Abbreviations: COVID-19: coronavirus disease 2019; IQR: interquartile range.

The proportion of patients who died was significantly different between clusters (P < .001, V = .092), with the “Decliners” containing a greater proportion of patients who died compared to the “Risers” and “Peakers” (10%, n = 233; 6%, n = 193; and 4%, n = 88, respectively). Moderate and critical COVID-19 severity were also significantly different between clusters (P = .027, V = .031 and P < .001, V = .059, respectively), with the “Peakers” containing the lowest proportion of moderate and critical cases (3%, n = 65 and 5% n = 119, respectively).

Glucose levels (P < .001 for all 3 comparisons) and the proportion of patients with hyperglycemic glucose measurements before, during, and after COVID-19 diagnosis (P < .001 for all 3 comparisons, V = .313, .208, and .309, respectively) were significantly different between clusters. The “Decliners” demonstrated the highest median and greatest proportion of hyperglycemia before COVID-19 diagnosis (median (interquartile range [IQR]) = 6.60 (1.83) mmol/L; 40%, n = 925). The “Peakers” demonstrated the highest median and greatest proportion of hyperglycemia during COVID-19 diagnosis (median (IQR) = 6.16 (1.39) mmol/L; 26%, n = 561), while the “Risers” demonstrated the highest median and greatest proportion of hyperglycemia after COVID-19 diagnosis (median (IQR) = 6.66 (1.94) mmol/L; 40%, n = 1191).

Radar visualization

Significantly different scaled categorical features were displayed in a radar diagram for each cluster and features with a value greater than 0.5 are reported. “Risers” were characterized by hyperglycemia after COVID-19 diagnosis and moderate and critical COVID-19 severity (Figure 1a, bottom). “Decliners” were characterized by the greatest proportion of patients who died and hyperglycemia before COVID-19 diagnosis (Figure 1B, bottom). “Peakers” were characterized by hyperglycemia during COVID-19 diagnosis (Figure 1C, bottom).

Nineteen clusters model

Cluster trajectories and centers from the 19 cluster model, segment trajectory dataset, and ts k-means model are depicted in Figure 2, from which 8 general patterns were visually identified according to cluster centers.

Figure 2.

Glucose trajectories of the 19 clusters from the segment and k-means model. (A) Glucose trajectory of the ‘Risers’ cluster with black lines representing individual trajectories and the red line representing cluster centers. (B) Glucose trajectory of the ‘Peakers’ cluster with black lines representing individual trajectories and the red line representing cluster centers. (C) Glucose trajectory of the ‘Decliners’ cluster with black lines representing individual trajectories and the red line representing cluster centers. (D) Glucose trajectory of the ‘Valleyers’ cluster with black lines representing individual trajectories and the red line representing cluster centers.

The first set of clusters were labeled as “Risers” because all cluster means demonstrated an overall increase in scaled glucose values and varied by timing of that increase (Figure 2A). Clusters 3 and 16 were included into the “Steady Risers” because their centers demonstrated a steady increase in scaled glucose during and after COVID-19 diagnosis. Clusters 6, 12, and 13 were included into the “Delayed Risers” because their centers demonstrated a delayed increase in scaled glucose that occurred after COVID-19 diagnosis. Clusters 7 and 11 were included into the “Early Risers” because their centers demonstrated an early increase in scaled glucose during and persisted after COVID-19 diagnosis. The next set of clusters were labeled as “Peakers” because their centers demonstrated a peak in scaled glucose during COVID-19 diagnosis that subsequently returned to baseline (Figure 2B). Clusters 2, 10, 14, and 15 were included into the “Peakers.”

The next set of clusters were labeled as “Decliners” because their centers demonstrated an overall decrease in scaled glucose values and varied by timing of that decrease (Figure 2C). Cluster 18 was included in “Steady Decliners” because its center demonstrated a steady decline in scaled glucose during and after COVID-19 diagnosis. Clusters 4 and 9 were included in “Delayed Decliners” because their centers demonstrated a delayed decline in scaled glucose after COVID-19 diagnosis. Clusters 5, 8, and 19 were included into the “Early Decliners” because their centers demonstrated an early decline in scaled glucose that occurred during and persisted following COVID-19 diagnosis. The final set of clusters were labeled as “Valleyers” because their centers demonstrated a drop in scaled glucose during COVID-19 diagnosis that returned to baseline afterward (Figure 2D). Clusters 1 and 17 were included into the “Valleyers.”

Summary statistics for all features were calculated for each cluster (Table 2).

Table 2.

Statistical summary of phenotypic features of 19 clusters from the segment and k-means model

Feature		Steady risers (n = 891)	Delayed risers (n = 1583)	Early risers (n = 690)	Steady decliners (n = 346)	Delayed decliners (n = 578)	Early decliners (n = 1189)	Peakers (n = 1392)	Valleyers (n = 833)	P-value
Demographics
Age	Median (IQR) (years)	58 (31)	59 (27)	58 (31)	54 (32)	56 (32)	58 (28)	58 (28)	62 (26)	P < .001
	% > 65 years (n)	36% (318)	36% (570)	34% (233)	31% (106)	31% (180)	35% (416)	33% (465)	40% (337)	P = .004
										V = .052
Sex	% Female (n)	59% (530)	54% (855)	56% (383)	60% (208)	60% (345)	55% (653)	56% (773)	54% (450)	P = .042
										V = .044
Death	Median (IQR) (years)	67 (22)	70 (25)	76 (21)	68 (16)	73 (24)	71 (25)	65 (23)	70 (16)	P = .578
	% Death (n)	4% (33)	6% (99)	3% (24)	6% (22)	5% (30)	11% (130)	4% (53)	15% (123)	P < .001
										V = .147
COVID-19 severity
Moderate	% (n)	4% (33)	5% (73)	2% (15)	2% (15)	3% (16)	3% (40)	3% (47)	6% (46)	P = .004
										V = .053
Severe	% (n)	4% (34)	4% (61)	4% (26)	3% (10)	3% (16)	3% (33)	3% (39)	3% (25)	P = .563
Critical	% (n)	8% (74)	10% (157)	5% (37)	4% (15)	4% (25)	7% (86)	6% (84)	12% (100)	P < .001
										V = .089
Medications
Anti-hyperglycemic agents or insulin	% (n)	7% (62)	7% (107)	5% (36)	6% (22)	5% (27)	6% (72)	7% (97)	9% (71)	P = .106
Glucose	% (n)	13% (113)	12% (185)	10% (69)	12% (42)	12% (70)	12% (139)	13% (177)	14% (120)	P = .321
Steroids	% (n)	16% (142)	15% (233)	13% (89)	16% (56)	12% (72)	14% (165)	14% (193)	16% (130)	P = .395
Glucose levels
Before	Median (IQR) (mmol/L)	5.16 (0.89)	5.72 (1.22)	5.05 (0.94)	6.33 (1.44)	5.94 (1.17)	6.88 (2.05)	5.39 (1.05)	6.55 (1.94)	P < .001
	% Hyperglycemic (n)	5% (44)	16% (261)	6% (42)	29% (102)	16% (94)	47% (559)	10% (141)	39% (328)	P < .001
										V = .370
During	Median (IQR) (mmol/L)	5.72 (1.00)	5.44 (1.00)	5.94 (1.17)	5.72 (0.89)	5.88 (1.11)	5.49 (1.00)	6.33 (1.61)	5.33 (1.00)	P < .001
	% Hyperglycemic (n)	11% (96)	8% (129)	19% (134)	12% (41)	15% (87)	8% (96)	31% (426)	8% (65)	P < .001
										V = .243
After	Median (IQR) (mmol/L)	6.27 (1.44)	6.88 (2.11)	5.83 (1.11)	5.22 (0.89)	5.11 (0.89)	5.77 (1.17)	5.38 (1.00)	6.60 (2.05)	P < .001
	% Hyperglycemic (n)	29% (259)	46% (736)	17% (115)	7% (25)	5% (30)	15% (175)	9% (132)	40% (333)	P < .001
										V = .359
New-onset diabetes diagnoses
Diabetes	% (n)	1% (10)	2% (25)	1% (4)	0% (1)	1% (3)	2% (22)	2% (23)	2% (15)	P = .048
										V = .043

Note: Statistical summary of all features is presented for each cluster. Differences in continuous variables were tested using Kruskal–Wallis H-test. Differences between clusters in categorical variables were tested using the Chi-squared test for independence with Yates Continuity Correction and Cramer’s V for effect size was calculated for statistically significant features. P-values were evaluated at the .05 significance level, and the bold text indicates statistical significance.

Abbreviations: COVID-19: coronavirus disease 2019; IQR: interquartile range.

Age (P < .001) and proportion of patients greater than 65 years (P = .004, V = .052) were significantly different among groups, with “Valleyers” having the oldest and greatest proportion of patients older than 65 (median (IQR) = 62 (26) years; 40%, n = 337). The proportion of females were also significantly different among groups, though the effect size was small (P = .042, V = .044). The proportion of patients who died was significantly differ among groups (P < .001, V = .147), with the greatest proportion of death observed for the “Valleyers” (15%, n = 123). Moderate and critical COVID-19 severity were significantly different among groups (P = .004, V = .053 and P < .001, V = .089, respectively). “Valleyers” and “Delayed Risers” demonstrated the greatest proportion of patients with moderate (6%, n = 46 and 5%, n = 73, respectively) and critical severity (12%, n = 100, and 10%, n = 157, respectively).

Glucose levels (P < .001 for all 3 comparisons) and the proportion of patients with hyperglycemia before, during, and after COVID-19 diagnosis (P < .001 for all 3 comparisons, V = .370, .242, and .359, respectively) were significantly different between groups. The greatest median glucose and highest proportion of hyperglycemia were attributed to “Early Decliners” before COVID-19 diagnosis (median (IQR) = 6.88 (2.05) mmol/L; 47%, n = 559, respectively), “Peakers” during COVID-19 diagnosis (median (IQR) = 6.33 (1.61) mmol/L; 31%, n = 426, respectively), and “Delayed Risers” after COVID-19 diagnosis (median (IQR) = 6.88 (2.11) mmol/L; 46%, n = 736, respectively). Diagnosis of diabetes mellitus was also significantly different between groups (P = .048, V = .043), with “Delayed Risers” containing the greatest number of patients with new-onset Diabetes (2%, n = 25).

Significantly different scaled categorical features were displayed in a radar diagram for each group (Figure 3) and features with a value greater than 0.5 are reported. “Steady Risers” were characterized by female sex and hyperglycemia after COVID-19 diagnosis, “Delayed Risers” were characterized by patients aged 65 or older, moderate and critical COVID-19 severity, hyperglycemia after COVID-19 diagnosis, and new-onset diabetes mellitus, and “Early Risers” were characterized by hyperglycemia during COVID-19 diagnosis (Figure 3A). “Peakers” were characterized by hyperglycemia during COVID-19 and new-onset diabetes mellitus (Figure 3B). “Steady Decliners” were characterized by female sex and hyperglycemia before COVID-19 diagnosis, “Delayed Decliners” were characterized by female sex, and “Early Decliners” were characterized by death, hyperglycemia before COVID-19 diagnosis, and new-onset diabetes mellitus (Figure 3C). “Valleyers” were characterized by patients aged 65 or older, death, moderate and critical COVID-19 severity, hyperglycemia before and after COVID-19, and new-onset diabetes mellitus (Figure 3D).

Figure 3.

Radar plot of phenotypic features of the 19 clusters from the segment and k-means model. (A) Radar plot of the scaled statistically significant features of the ‘Risers’ cluster. (B) Radar plot of the scaled statistically significant features of the ‘Peakers’ cluster. (C) Radar plot of the scaled statistically significant features of the ‘Decliners’ cluster. (D) Radar plot of the scaled statistically significant features of the ‘Valleyers’ cluster.

DISCUSSION

This study provides the first data-driven assessment of random blood glucose measurements in a large, multicenter cohort of patients diagnosed with COVID-19 in the United States. Time-series clustering identified 2 optimal models for characterizing glucose trajectories: a low-complexity model with 3 clusters and a high-complexity model with 19 clusters. Cluster membership for both models was associated with proportions of patients with moderate and critical COVID-19 severity, mortality, and hyperglycemia at each timepoint, and mean glucose values at each timepoint (Tables 1 and 2). The 19 cluster model was additionally associated with the proportion of patients with age older than 65 years, female sex, and new-onset diabetes mellitus diagnosis, and mean age at death (Table 2). These results highlight the utility of using unsupervised clustering to learn novel patterns of glucose changes and call for longitudinal glucose monitoring in patients with COVID-19.

The simpler 3 cluster model consisted of “Risers,” “Decliners,” and “Peakers.” Findings from the “Risers” suggest that there is a group of individuals who experience moderate or critical COVID-19 with associated increases in blood glucose after resolution of disease. Though this does not indicate an increase in new-onset diabetes mellitus, this work shows that subclinical glucose dysfunction can occur in patients following COVID-19. The “Decliners” demonstrated that elevated glucose before COVID-19 diagnosis is associated with death. These findings support previous evidence that hyperglycemia before COVID-19 indicates a poor prognosis., Finally, findings from the “Peakers” suggest that a group of patients experience elevated blood glucose during COVID-19 diagnosis that is not characterized by moderate and critical COVID-19 severity, supporting previous evidence that coronaviruses can induce transient hyperglycemia. The 19 cluster model consisted of 8 groups that captured a wide variety of subtle changes in glucose trajectories that remained masked in the 3 cluster model. The “Early Decliners,” “Valleyers,” “Delayed Risers,” and “Peakers” were characterized by a sharp peak in diabetes diagnosis. Of these 4 groups, only the “During Decliners” and “Delayed Risers” were associated with moderate or critical COVID-19 severity. Together, these findings suggest that glucose trajectories may be associated with increased diabetes incidence following SARS-CoV-2 infection in both symptomatic and asymptomatic patients.

This work highlights the potential for data-driven techniques to elucidate longitudinal glycemic dysfunction in patients with COVID-19. Unlike previous studies characterizing glucose in COVID-19,,,, this work utilizes electronic medical records from a large cohort of patients across the United States. Presentation of a low- and high-complexity model emphasizes the heterogeneity in glucose curves of patients with COVID-19. The presented analytical approach can serve as a framework for future data-driven analyzes in longitudinal glycemic evaluation of COVID-19 patients. In addition, the presented analysis demonstrates the utility of real-world clinical data in longitudinal research analyses. This work also captures COVID-19 cases from a broader range of pandemic, as previous studies are limited by short time frames that often coincide with peak periods of the pandemic., There is a dearth of knowledge on the long-term outcomes of patients with COVID-19 because longitudinal studies of the clinical sequela are limited, and this study contributes to the longitudinal understanding of glycemic changes following COVID-19 diagnosis.

To the best of our knowledge, this is the first work assessing the role of COVID-19 in the etiology of diabetes using real-world clinical data. Many viruses, including enterovirus, Coxsackievirus B, rotavirus, mumps virus, cytomegalovirus, and rubella, have been associated with diabetes pathogenesis in epidemiological studies., Similarly, these results provide hypotheses for further investigation into why certain individuals may develop diabetes following a COVID-19 infection. Though this work follows individuals for a short period of time surrounding COVID-19 infection, these results highlight the need for longer-term follow-ups and the development of trajectories for other diabetes-related parameters, as evidenced by previous studies investigating body mass index and glycosylated hemoglobin. Initiatives such as the National COVID Cohort Collaborative (N3C) provide opportunities for addresses longer-term follow-up and diabetes-related clinical consequences of COVID-19. N3C leverages the vast amount of electronic medical record collected through the course of the pandemic and aims to facilitate novel scientific discoveries to aid in responding to the pandemic.

There are several limitations of this work. First, it is possible that fewer individuals with asymptomatic or mild disease sought medical care at large healthcare organizations that are represented by TriNetX. Due to inherent limitations in electronic medical record data, distinguishing asymptomatic and mild COVID-19 was not possible. In addition, recent work suggests that asymptomatic COVID-19 cases are more prevalent than previously thought. Therefore, these results may represent glycemic changes in a more severe spectrum of COVID-19 cases. Next, this study only included individuals with glucose measurements before, during, and after COVID-19, resulting in a significant reduction of patients from the original cohort (Supplementary Figure S1). Duration of COVID-19 is highly variable in patients and our stratification may oversimply the COVID-19 time course. However, this stratification of glucose timepoints was necessary because electronic medical record data are highly variable and identifying a consistent time axis for analysis was not possible. This was partially accounted for by comparing the performance of multiple time intervals, and our findings call for longitudinal cohort studies in patients diagnosed with COVID-19 with consistently and uniformly collected glucose measurements. In addition, these inclusion criteria may favor patients who receive care continuously from a single medical center. Further work is needed to validate the impact of COVID-19 on longitudinal glucose changes in populations who do not receive consistent care. Additionally, there were a variable number of glucose measurements for each patient included in generating the segment trajectory dataset (Supplementary Methods). A subset of patients only contained single random glucose levels per timepoint, which may under-represent overall glucose levels for the given patients. Of the 7502 patients included in this study, 7230 patients had more than 1 glucose measurement per timepoint, suggesting that the bias due to the number of random blood glucose measurements may be minimal. Given that the average time to COVID-19 symptom onset is 5 days and viral detection 2 weeks after symptom onset was minimal, we defined 3 weeks to encompass COVID-19 duration. However, the time to recovery for COVID-19 patients is variable and more robust methods may be needed to capture disease duration. Random blood glucose measurements were selected as they are appropriate estimates for glycemic control in noninsulin-dependent patients. However, their use is imprecise, and further studies should be conducted using fasting blood glucose measurements in well-defined cohorts recruited for specific studies. While our cluster groups were derived using subjective methods, we plan to validate these phenotypes using external data-centric methods with availability of longer post-COVID records for these patients in TriNetX and N3C.

CONCLUSIONS

This work utilizes time-series unsupervised clustering techniques to identify distinct groups of glucose trajectories. We provide a detailed analysis of 2 models corresponding to 3 and 19 clusters. Both models provide insights on how cluster features can be associated with clinical phenotypes and expected outcomes. This work demonstrates that unsupervised learning models of different complexities can be used to provide a clinically useful stratification of glucose trajectories in patients diagnosed with COVID-19. Further subtyping and longer follow-up can identify subphenotypes of patients who are more susceptible to glycemic dysregulation following COVID-19 infection and inform appropriate point-of-care guidelines for COVID-19 management.

FUNDING

This work was supported by the National Library of Medicine (T15LM5920634 to SM) and the National Center for Advancing Translational Sciences (UL1TR002538). Computational resources were provided by the University of Utah Center for High Performance Computing, partially funded by the NIH Shared Instrumentation Grant (1S10OD021644-01A1). De-identified data were provided by TriNetX, a global health research network compliant with Section 164.514(b)(1) of the HIPAA Privacy Rule.

Click here for additional data file.