Sandya Subramanian

The paper applies a lognormal mixture neural temporal point process to model heartbeat dynamics, showing improved robustness to noise compared to an inverse Gaussian model.

The paper evaluates scaling laws for stochastic sequence data using distributional calibration metrics, showing that performance improves with cohort size beyond point accuracy saturation.

This study evaluates how well Time-Series Foundation Models predict heartbeat dynamics and finds they fail to capture long-range temporal dependencies, leading to increasing prediction errors over time.

This paper develops a dilated CNN model using HR-EGG and accelerometry data to detect meals passively in ambulatory settings, achieving high accuracy.

This paper develops an unsupervised machine learning framework to remove surgical artifacts from EDA data in clinical settings, achieving high sensitivity and specificity.

This study assesses the relationship between behavioral and autonomic dynamics during propofol-induced unconsciousness using statistical models and physiological data from nine volunteers.

The paper models EDA pulse amplitudes as differences of inverse Gaussian distributions, showing they outperform other models in describing physiological sweat production.

The paper analyzes electrodermal activity (EDA) using probability models to characterize its temporal structure and validate them with experimental data.

The paper develops a network analysis algorithm to localize the epileptogenic zone using invasive EEG recordings, showing better agreement with clinical outcomes in successful surgeries.