Case-based reasoning using electronic health records efficiently identifies eligible patients for clinical trials

Abstract

Objective: To develop a cost-effective, case-based reasoning framework for clinical research eligibility screening by only reusing the electronic health records (EHRs) of minimal enrolled participants to represent the target patient for each trial under consideration.

Materials and methods: The EHR data--specifically diagnosis, medications, laboratory results, and clinical notes--of known clinical trial participants were aggregated to profile the "target patient" for a trial, which was used to discover new eligible patients for that trial. The EHR data of unseen patients were matched to this "target patient" to determine their relevance to the trial; the higher the relevance, the more likely the patient was eligible. Relevance scores were a weighted linear combination of cosine similarities computed over individual EHR data types. For evaluation, we identified 262 participants of 13 diversified clinical trials conducted at Columbia University as our gold standard. We ran a 2-fold cross validation with half of the participants used for training and the other half used for testing along with other 30 000 patients selected at random from our clinical database. We performed binary classification and ranking experiments.

Results: The overall area under the ROC curve for classification was 0.95, enabling the highlight of eligible patients with good precision. Ranking showed satisfactory results especially at the top of the recommended list, with each trial having at least one eligible patient in the top five positions.

Conclusions: This relevance-based method can potentially be used to identify eligible patients for clinical trials by processing patient EHR data alone without parsing free-text eligibility criteria, and shows promise of efficient "case-based reasoning" modeled only on minimal trial participants.

Keywords: artificial intelligence; clinical trials; electronic health records; information storage and retrieval.

Figure 1:

Overview of the “case-based reasoning” framework to discover eligible patients for a clinical trial through the “target patient,” a representation of the trial derived from the EHR data of a minimal sample of participants.

Figure 2:

Overview of the process to derive the clinical trial’s “target patient” by modeling the EHR data of minimal enrolled participants.

Figure 3:

Classification results in terms of the area under the ROC curve averaged over both the evaluation folds. A patient was considered eligible if its relevance score with the corresponding “target patient” was over a threshold (ranged between 0 and 1), ineligible otherwise.

Figure 4:

Precision-at-5 (P5) obtained by w-comb for every fold and every clinical trial. We derived each trial’s “target patient” from the corresponding training participants, ranked the test patients by their relevance with the trial, and measured how many eligible patients were within the top five positions. Results for upper-bound are included for comparison as well; in this case, there is no distinction between folds because results were identical.