The Work
5 verticals · 13 active · 13 past
Five research verticals span the full translational pipeline, from preclinical neural decoding to hospital-wide operational AI.
01 / Point-of-care AI
Develop clinical decision support tools to diagnose and predict clinical outcomes and trajectories.
02 / Operational AI
Develop operational decision support tools to provide enterprise insights and assist in health system wide issues.
03 / Anatomical Data AI
Develop accurate in-silico models of human anatomy using multimodal imaging and AI.
04 / Autonomic Nervous System AI
Develop algorithms that use non-invasive physiological data to diagnose disease presence & severity and predict treatment efficacy.
05 / Preclinical AI
Develop algorithms that use neural recordings from preclinical models to diagnose & predict disease states.
Past projects
This project focuses on developing generative deep learning models to reconstruct multi-lead electrocardiogram signals from a limited set of input ECG leads. By learning the relationships among ECG leads, the model aims to evaluate how accurately clinically meaningful cardiac signal information can be recovered from minimal-lead recordings. The project assesses reconstruction performance from both a signal-fidelity perspective and a clinical relevance perspective, including the preservation of diagnostically important ECG patterns. More broadly, this work explores whether a small number of key ECG leads can capture sufficient information to support robust ECG interpretation, scalable monitoring, and future AI-enabled clinical applications.
Predicting combined deterioration outcomes from 18 EHR features. Trained on over 1.5 million hospitalizations, the model outperforms the Epic Deterioration Index by 25% and is currently silently running in 7 hospitals, with plans to go live across all Epic Northwell hospitals.
Deep learning on wearable physiological data predicts clinical deterioration up to 17 hours in advance with greater than 94% sensitivity. Funded by a 4-year, $3.2M NIH grant. The algorithm is wearable-agnostic, validated across VitalConnect, VitalPatch, Biobeat, and Isansys devices.
This project focuses on developing machine learning models to support early risk stratification of patients presenting with acute coronary syndrome using electrocardiographic features, vital signs, and demographic information available at the point of first medical contact. Using data from a large, diverse healthcare system, the model evaluates whether quantitative ECG features extracted from admission electrocardiograms can predict in-hospital mortality and identify patients at elevated risk for clinical deterioration. The project also uses model interpretability methods to assess the relative contribution of ECG-derived variables compared with conventional clinical features. More broadly, this work explores how routinely acquired ECG data can be leveraged as an AI-enabled prognostic tool to support timely triage, escalation of care, and clinical decision-making in acute cardiovascular settings.
Development and validation of an LLM-powered framework for automated annotation of clinical charts as positive or negative for delirium, coupled with a supervised binary classification model for early prediction of delirium onset based on clinical and physiological data.
Continuous vital sign and heart rate variability monitoring during labor predicts maternal fever 2-3 hours before onset — a key risk factor for neonatal early-onset sepsis. HRV is one of the strongest predictors.
DeepAR forecasting models predict nursing demand and attrition up to 12 months ahead, using historical demand and attrition data combined with demographic, role-related, and job-related features. Preemptive hiring strategy is estimated to save close to $10 million annually across just 10 of Northwell's 400+ units. Ongoing work clusters separation phenotypes to identify modifiable factors.
60 vagus nerves (30 left, 30 right), over 3 million micro-CT images, and roughly 200 terabytes of data. 3D nnU-Net segmentation builds the most detailed human vagus digital twin to inform the design of selective vagus nerve stimulation therapies. $6.7M NIH grant (2023–25) with the TNP Lab.
Predicting which drug-resistant epilepsy patients will respond to vagus nerve stimulation therapy. Current clinical success rate: 65% of patients achieve more than 50% seizure reduction.
Machine learning on autonomic biomarkers (heart rate variability and related signals) to detect PTSD presence and severity from non-invasive physiological data. Trained and tested on 152 sessions from 30 PTSD patients and 49 healthy controls. Limited current treatment options motivate development of vagus nerve stimulation as a viable alternative.
Decoding vagus nerve signals related to inflammation in preclinical models, building on the 2018 PNAS finding that cytokine-specific neural activity can be decoded from the cervical vagus nerve.
Decoding vagus nerve signals related to metabolic states in preclinical models, including identification of hypoglycemia-specific neural signals.
Chronic vagus nerve recording in implanted mouse models to predict disease severity and treatment efficacy in inflammation models (CIA, CAIA), with longitudinal compound action potential tracking up to 6 months.