← Back to Research

Interpretability Research Team

6 Team Members Founded 2024

Our interpretability team focuses on understanding how AI models and biosensing systems work internally. By developing novel methods to interpret complex systems, we create a foundation for safety, trustworthiness, and clinical adoption of AI in healthcare.

Mission

We believe that trustworthy AI in healthcare requires complete transparency. Our mission is to develop interpretability techniques that allow clinicians and patients to understand exactly why AI systems make specific recommendations or predictions. This understanding enables safer deployment and higher confidence in AI-assisted medical decision-making.

Research Focus Areas

Time Series Interpretability

Developing specialized interpretability methods for understanding deep learning models applied to continuous health monitoring data. This includes attention visualization, temporal feature importance, and pattern discovery in biosensing time series.

SHAP and Attribution Methods

Adapting and extending SHAP values and gradient-based attribution techniques for time-series health data. These methods identify which biomarker patterns and temporal features drive individual predictions.

Neuron Analysis and Concept Learning

Investigating what individual neurons learn to represent in models trained on health data. By understanding learned concepts, we can identify clinical insights encoded in the models.

Counterfactual Explanations

Generating realistic "what-if" scenarios that explain model decisions. This helps clinicians understand what changes in biomarkers or behaviors would alter an AI recommendation.

Current Projects

Interpretable Anomaly Detection

Deep learning models for detecting anomalies in biosensing time series with built-in interpretability through attention mechanisms.

Model-Agnostic Explanations

Developing post-hoc explanation techniques that work with any health prediction model without requiring model retraining.

Clinical Validation Studies

Running studies with clinicians to validate whether our interpretability methods actually improve clinical understanding and decision-making.

Interactive Visualization Tools

Building web-based tools that let clinicians explore model explanations and understand AI predictions intuitively.

Key Publications

Our interpretability research has been published in top-tier venues including NeurIPS, ICML, and clinical AI conferences. Key papers focus on understanding neural patterns in health prediction, SHAP-based attribution for biosensing, and clinical validation of ML interpretability.

Team Values

Join Our Team

We're hiring researchers, engineers, and clinicians who are passionate about making AI interpretable and trustworthy. Interested in joining the interpretability team? Explore open positions on our careers page.