Objective:

Predict neurologic outcomes for comatose patients following cardiac arrest using natural language processing from electronic health records (EHR).

Background:

Clinical notes from the EHR are an abundant and heterogeneous source that captures a patient's clinical course. Machine learning algorithms have the potential to identify patterns from clinical notes that may support prognostication following cardiac arrest.

Design/Methods:

The cohort included comatose adult cardiac arrest patients from two hospitals in San Francisco and the MIMIC-III database. Patient neurologic outcomes were classified as Good (CPC 1-3) or Poor (CPC 4-5) at hospital discharge based on manual review of clinical notes. Clinical notes were reduced to 200 characters following specific headers, concatenated, tokenized, and lemmatized to extract n-gram features. A logistic regression model with L-BFGS optimization was trained on 70% of patients to predict neurological outcomes at discharge. Model performance was evaluated on the remaining 30% of patients using AUROC and AUPRC scores. Performance on the model was then compared between predictions on all notes vs a subset of notes limited to 24 hours after cardiac arrest.

Results:

We gathered 4,988 notes from 357 patients, with 242 of these patients having poor outcomes at discharge. The most common note types included consult notes (33.6%) and nursing notes (55.9%). The model’s AUROC and AUPRC micro-averages were 0.9 and 0.89, respectively. The model predicted 80% of individuals in the poor outcome group (CPC 4-5) and 76% in the good outcome group (CPC 1-3) when using all notes. The model performance decreased 7% when comparing predictions from all notes to 24 hours of notes.

Conclusions:

Our model had excellent performance for discriminating good and poor outcomes but had moderate accuracy. The model changed prediction class over time in 7% of patients, highlighting the dynamic process in the prognostic assessment of neurological recovery by clinicians in cardiac arrest.