Learning to predict post-hospitalization VTE risk from EHR data

Abstract

We consider the task of predicting which patients are most at risk for post-hospitalization venothromboembolism (VTE) using information automatically elicited from an EHR. Given a set of cases and controls, we use machine-learning methods to induce models for making these predictions. Our empirical evaluation of this approach offers a number of interesting and important conclusions. We identify several risk factors for VTE that were not previously recognized. We show that machine-learning methods are able to induce models that identify high-risk patients with accuracy that exceeds previously developed scoring models for VTE. Additionally, we show that, even without having prior knowledge about relevant risk factors, we are able to learn accurate models for this task.

Figure 1.

A graphical overview of the data-extraction process. From the data warehouse, which contains six types of records, and the genetic data, we extract histories for each patient. We represent each patient history using a vector of binary variables. Pairs of variables are used to represent recent and not-necessarily-recent diagnoses, procedures, etc. We vary the representations considered by optionally (i) using a set of curated risk factors to limit the variables included in each patient vector, and (ii) adding the variables that represent the genetic profile of each patient.

Figure 2.

Survival curves for four relevant individual variables: E.coli infection, hypovolemia, levofloxacin use, and pathologic fracture of vertebrae. The solid curve shown at the top of the figure represents the survival curve for the entire patient cohort.

Figure 3.

Precision-recall curves for learned models on the curated representation.

Figure 4.

Survival curves for learned models on the curated representation.

Figure 5.

Precision-recall curves comparing models learned using the curated variable set and the unabridged variable set.

Figure 6.

Survival curves comparing models learned using the curated variable set and the unabridged variable set.

Figure 7.

Precision-recall curves comparing the curated variable set, the genetic variable set, and the combined curated-genetic variable set using the random forest learning algorithm.

Figure 8.

Precision-recall curves comparing models learned using the curated and unabridged variable sets to the Sanofi and Chicago risk assessment questionnaires.

Figure 9.

Survival curves comparing models learned using the curated and unabridged variable sets to the Sanofi and Chicago risk assessment questionnaires.