I work with Dr. Kee B. Park on passive, phone-based detection of early neurological impairment. The current MVP is a pair of screening models for stroke signs — dysarthria heard in speech, facial palsy read from images — trained on locally held and newly gathered public data.

• The speech model reduces each clip to 240 MFCC-derived features and classifies them with a small multilayer perceptron, trained on the TORGO and Italian EasyCall corpora with every split made at the speaker level — an earlier notebook split by clip, and its 99% accuracy was the model recognising voices, not dysarthria. Held-out TORGO speakers come in at 82% clip accuracy (AUC 0.88); the deployable artifact, a combined-corpus model quantized to a 110 KB TFLite file, reaches AUC 0.78.
• The face model gave up on raw pixels after a MobileNetV2 CNN scored AUC 0.96 internally and chance on unseen-source faces — dataset fingerprints, not clinical signs. It now measures 40 geometric asymmetries from FaceMesh landmarks under gradient boosting: 81% accuracy (AUC 0.93) on curated tests, and genuinely above chance across corpora (AUC 0.71 with per-video voting).
• The argument for the approach, co-authored with Park: AANS Neurosurgeon, November 2024.
• What the cross-corpus numbers make evident: a wider and more robust dataset is required for anything further — stronger generalisation, stroke-specific claims, and eventually clinical validation. The corpora at hand are off-etiology and a few dozen identities deep, and they carry the models only so far.

Under Dr. Robert David Stevens, I build a preoperative triage model that predicts which elective craniotomy patients will need ICU-level care. It is developed and validated on two cohorts — Johns Hopkins PMAP (n=310) and MIMIC-IV (n=367) — and the external validation runs both ways: train on one, test on the other.

• The rebuilt pipeline is repeated 10-fold nested cross-validation over four arms — a penalised logistic regression, XGBoost, CatBoost, and a stack of the three under a ridge meta-learner. Everything that could peek — imputation, scaling, correlation filtering, feature selection, tuning, calibration — fits inside the training folds, cross-cohort feature selection never sees a label, and the confidence intervals are bootstrapped at the patient level.
• The rebuild caught what the originally published analysis missed: seventeen ICU-stay variables had leaked into the MIMIC feature set, carrying roughly 70% of that model's SHAP importance. Removing them costs XGBoost 0.146 AUROC. MIMIC's published advantage was measurement timing, not a richer feature space.
• On clean features the two cohorts perform comparably, PMAP improved under the corrected pipeline (CatBoost, 0.729 to 0.774), and going fully preoperative costs nothing measurable. Every arm now beats admitting everyone to the ICU, in both transfer directions.
• One caveat, on the outcome itself: the composite leans on its softest component. Depressed consciousness is the sole trigger in 41% of events; among them are only four reoperations and one ICP crisis.