I work with Dr. Kee B. Park on passive, phone-based detection of early neurological impairment, and in practice that's training multimodal deep-learning classifiers on audio and image samples to pick up neuromotor and speech biomarkers.

• Movement and speech start to go before anyone gets a diagnosis: tremor, bradykinesia, and postural instability on the motor side, dysarthria and falling verbal fluency in speech. Facial palsy shows up in the image. The conditions are Alzheimer's, Parkinson's, and ischemic stroke.
• A model that only works on the population it was trained on is not the point. The models have to hold up across countries and health systems, and in the long run they need to work as well in a low-resource setting as in a high one.
• The argument for the approach, co-authored with Park: AANS Neurosurgeon, November 2024.

Under Dr. Robert David Stevens, I work on machine-learning models that stratify the risk of complications after an elective craniotomy. I develop them on MIMIC-IV and evaluate them on a Johns Hopkins PMAP cohort.

• The models put patients in order of risk, and that's all they do. They don't diagnose anyone, and the ranking says nothing about which complication or when.
• Somebody has to decide how closely a patient is watched once they leave the operating room, and that's the one decision the ranking is for.
• Each model refits its own feature selection inside every training fold — a penalised logistic regression, XGBoost, CatBoost, and a stack of the three under a ridge meta-learner. The second cohort is for testing the models, never for training them. Nobody knows yet whether the ranking still means anything when the patients, the surgeons, and the recording habits change.