Unsupervised Anomaly Detection to Characterize Heterogeneity in Type 2 Diabetes

Abstract

Diabetes is associated with heterogeneous behaviors affecting patients' clinical characteristics and trajectories. This study includes 21,288 patients with type 2 diabetes (women, ages 30 to 65). The cohort was filtered through a set of preprocessing heuristics in order to assure the cohort exhibited a similar clinical trajectory. Anomalous characteristics were then identified using dimensionality reduction and anomaly detection methods. Compared to the majority of the cohort, patients classified as anomalous were twice as likely to be admitted into the hospital (7.94[7.59 8.28] versus 3.12[3.06 3.17] times), have a higher incidence of comorbidities (2[1.64 2.36] times more), and be prescribed more insulin and less new and more expensive diabetes medications (such as Sodium glucose co-transporter 2 inhibitors). Patients with these anomalous characteristics may benefit from additional or specialized interventions to avert their risk for adverse outcomes.

Figure 1:

Cohort selection workflow. The steps to reach this cohort include: 1) filter for women, ages 30 to 65 who have at least two occurrences of type 2 diabetes, 2) filter for members with a quiescent period of at least 1 year of data without any record of type 2 diabetes, followed by 1 to 3 years (inclusive) of data while having type 2 diabetes, 3) extract the 1 to 3 years of data while having type 2 diabetes, 4) filter for those enrolled in a medical plan and have pharmacy benefits.

Figure 2:

Distribution and percent frequency of comorbidities in typical and outlier patients. Outlier patients have a higher incidence of comorbidities (2[1.64 2.36] times more) compared with that of the typical patients. (a) Shows the percent frequency of the most commonly found comorbidities in typical and outlier patients. (b) Shows the distribution of the number of comorbidities found in typical and outlier patients.

Figure 3:

Plots showing medication use across typical and outlier patients. Plots show the average percent frequency across the 3 different anomaly detection models, and show the progression over the years after type 2 diabetes appearance. The plots indicate usage across the cohort for (a) insulin, (b) Metformin, (c) Glyburide, (d) SGLT2is, (e) GLP-1, (f) DPP-4, and (g) PPAR-γ. Note there are some patients taking these medications before the appearance of type 2 diabetes in their records. This shows a possibility of miscoding for some patients, though there is a clear increase in intake once type 2 diabetes appears on the patients’ records.

Figure S1:

Distribution of anomaly scores across the three anomaly detection models. The threshold in each model was defined in order to account for the ratio of outlier to typical to be 10:90 respectively.