Scalable Electronic Phenotyping For Studying Patient Comorbidities

Abstract

Over 75 million Americans have multiple concurrent chronic conditions and medical decision making for these patients is mostly based on retrospective cohort studies. Current methods to generate cohorts of patients with comorbidities are neither scalable nor generalizable. We propose a supervised machine learning algorithm for learning comorbidity phenotypes without requiring manually created training sets. First, we generated myocardial infarction (MI) and type-2 diabetes (T2DM) patient cohorts using ICD9-based imperfectly labeled samples upon which LASSO logistic regression models were trained. Second, we assessed the effects of training sample size, inclusion of physician input, and inclusion of clinical text features on model performance. Using ICD9 codes as our labeling heuristic, we achieved comparable performance to models created using keywords as labeling heuristic. We found that expert input and higher training sample sizes could compensate for the lack of clinical text derived features. However, our best performing model included clinical text as features with a large training sample size.

Figure 1.

Full Model Framework. ICD9 codes were used to generate noisy labels, which were used as training examples in a LASSO logistic regression. The model labeled each patient as a case or control, which was validated against a gold standard clinician-reviewed dataset.

Figure 2.

Schematic showing the range of trade-offs that were explored to maximize model performance while minimizing model complexity.

Figure 3.

Model accuracy as a function of increasing training sample size. Red corresponds to performance of the model with clinical text as features, and blue corresponds to performance of the model without clinical text as features.

Figure 4.

Model iteration using feature anchoring and feature removal

Figure 5.

Models without clinical text were evaluated to assess whether increasing the training sample size would allow for comparable performance in models with and without clinical text as features. When anchoring was added to the model with large training sample size (N=5,000) without clinical text, it could perform as well as a model using clinical text as features.