Date of Award

January 2026

Document Type

Open Access Thesis

Degree Name

Medical Doctor (MD)

Department

Medicine

First Advisor

Nils Petersen

Abstract

Background - Current methods of identifying patients in the neurological intensive care unit (NeuroICU) with disrupted cerebral autoregulation rely on statistical methods that require significant time to produce predictions. There is a need for faster algorithms that can minimize the secondary damage caused by disrupted autoregulation in these patients.

Research Aims - We aim to develop a machine learning (ML) model that determines whether a patient is outside optimal autoregulatory limits using continuously recordable and minimally invasive physiologic data gathered in the NeuroICU.

Hypothesis - We hypothesize we can use ML techinques to ingest physiologic data in short time windows and determine the autoregulatory status of a patient in a generalizable manner, performing these predictions in a shorter time window than the pre-existing algorithms.

Methods - Physiologic data was collected between 2017 – 2022 from 210 patients admitted to the Yale-New Haven Hospital Neurointensive Care Unit. All patients were admitted with a diagnosis of acute large vessel occlusion stroke and received thrombectomy. We train approximately 5,200 models on windows extracted from this data to predict whether the patient is outside autoregulation limits previously calculated using established algorithms. We perform a hyperparameter search on each of these models. We extract the best performing models in terms of area under the receiver operating characteristic curve (AUROC) and balanced accuracy.

Results - A Random Forest model trained on 30 minute windows achieves an AUROC of 0.67 (95% CI: 0.65 – 0.69) and balanced accuracy of 64% (95% CI: 62% – 66%). This is equivalent to approximately 1.7 true alerts for every 1 false alert. Models using windowsof up to 5 minutes achieve similar accuracy.

Impact - While not achieving clinically relevant accuracy, we offer a proof of concept for ML approaches on neurophysiologic data. However, our findings should be interpreted cautiously. Although performance may improve with additional features or reduced noise, it remains possible that an alternative analytic framework will be required.

Comments

This is an Open Access Thesis.

Download
Open Access

This Article is Open Access

Share

COinS