A log of all my publications
This study introduces EarCO, a non-invasive cardiac output monitoring system using commodity earbuds with photoplethysmography and ballistocardiogram signals. The research demonstrates that multimodal earbud sensors can accurately estimate cardiac output for daily cardiovascular health monitoring.
This study shows that morphological features from earbud PPG sensors outperform heart rate variability features for stress detection. The research evaluates earbud sensors as a platform for stress monitoring.
A comprehensive Python package and web interface that implements computational models of memory search for analyzing verbal fluency task data. The tool provides multiple automated methods for cluster-switch analysis and foraging models to study semantic memory search patterns in both research and clinical settings.
This study investigates the link between speech signals and psychological distress symptoms using the Distress Analysis Interview Corpus. The research demonstrates that speech behavioral markers align primarily with somatic and affective symptom factors rather than cognitive alterations.
A novel predictive model called 'Semantic Scent' that extends optimal foraging theory to predict when switches occur during semantic memory search. The model outperforms existing approaches by accurately predicting cluster transitions based on the proximity of yet-to-be-produced items in semantic space.
This study validates speech-based markers for cognitive impairment detection by comparing original and expanded linguistic feature sets against neuropsychological tests. The research demonstrates that expanded linguistic features outperform traditional screening tools like MMSE in classifying cognitive status.
This study develops multimodal machine learning models to automatically detect client behavioral codes in motivational interviewing sessions using BERT and VGGish encoders. The research achieves an F1-score of 0.72 for three-class classification of client utterances and explores associations between in-session language patterns and subsequent behavioral outcomes.
This study demonstrates that voice and language features can predict cognitive impairment with high accuracy (AUC 0.942) using data from the Framingham Heart Study. The research identifies specific acoustic and linguistic biomarkers that correlate with neuropsychological test results and dementia diagnoses.
This study explores and validates voice features extracted from recorded speech samples as digital biomarkers for depression symptoms including suicidality, psychomotor disturbance, and depression severity. Voice features from depressed subjects successfully predicted PHQ9 scores with an area under the curve of 0.821 and mean absolute error of 4.7.
A comprehensive analysis of the AVEC 2019 Detecting Depression with AI Sub-Challenge dataset, examining acoustic and linguistic digital biomarkers for depression detection. The study identifies key preprocessing considerations and modeling methodologies while highlighting important dataset limitations and bias issues that affect model generalizability.