Research

• Research
  • Bio-signal processing
  • Interpretable machine learning
  • Disease prediction
  • Patient similarity
  • Neural engineering

Bio-signal processing

I specialize in deep learning algorithms for physiological signals collected from wearable devices, including electrocardiograms (ECG), photoplethysmography (PPG), phonocardiograms (PCG, heart sounds), and arterial blood pressure (ABP). Many of these solutions have been productionized and ranked first place at the PhysioNet Challenges.

• 2021 – Interpretable Additive Recurrent Neural Networks For Multivariate Clinical Time Series
• 2020 – A Wide and Deep Transformer Neural Network for 12-Lead ECG Classification
• 2019 – A Multi-Task Imputation and Classification Neural Architecture for Early Prediction of Sepsis from Multivariate Clinical Time Series
• 2018 – Densely connected convolutional networks for detection of atrial fibrillation from short single-lead ECG recordings
• 2018 – Analyzing single-lead short ECG recordings using dense convolutional neural networks and feature-based post-processing to detect atrial fibrillation
• 2016 – Ensemble of feature-based and deep learning-based classifiers for detection of abnormal heart sounds

Interpretable machine learning

A key requirement of clinical models is that they have to be explainable so that the nurse or clinician can understand the risk factors. The models should also be editable so we can identify when then algorithm makes an error and fix it before deployment. I created the Interpretable-RNN, which is an additive sequence learning model that achieves both high accuracy and is fully glassbox.

I developed the neural decision tree model which is a neural representation of highly interpretable gradient boosted decision trees. This allowed us to scale an interpretable class of models in a federated privacy preserving architecture.

Disease prediction

I built realtime clinical risk algorithms that are deployed on patient monitors and central stations.

Early warning systems like the Hemodynamic Stability Index identifies ICU patients that will need fluids or pressors in the near future. The HSI model is designed for Philips bedside patient monitors and is the flagship machine learning model in hemodynamics analytics products. This work was extended with federated learning, a privacy preserving technique that was used to train the model across hundreds of ICUs.

Machine learning can also be used to improve clinical trial design by finding the right patient at the right time. I led a collaboration with Bayer Pharmaceuticals to optimize clinical trials investigating drugs to treat respiratory distress.

Patient similarity

Patient similarity aims to find cohorts based on vital signs, labs, medical history, and treatments for case-based reasoning. How can we learn useful similarity functions that compare physiological state between patients? We developed a similarity metric to find patients-like-me in electronic health records.

Prior medical knowledge is widely available in electronic health records but are not utilized because of practical constraints on data availability or cost of acquiring the data to make inferences. I developed a a novel approach to learn from metadata called prior knowledge-guided learning using a mixture-of-experts model where gating probabilities are tuned by a nearest neighbors graph adjacency matrix.

Neural engineering

My research in neural engineering spans computational modeling, animal models, and the analysis of brain signals like evoked potentials, EMG, EEG, and EOG.