LLMs vs. traditional ML
Compared LLM based semantic models against XGBoost across clinical prediction tasks.
Result. Text representation critically affects
performance. Traditional ML still wins on pure physiological signals.
Seven pioneering initiatives across clinical ML, NLP, computer vision, and causal inference.
Below is a selection of research led by our inaugural cohort of 2025.
Compared LLM based semantic models against XGBoost across clinical prediction tasks.
Built an ETL first prototype for clinical evidence synthesis using a Knowledge Graph RAG system.
Used SEER registry data linked with social determinants of health proxies to stratify prostate cancer patients.
Investigated whether a simplified multimodal model using only PET/CT and clinical data could match expert annotated performance.
Built a training simulator using LLMs to create interactive, data grounded emergency triage encounters from MIMIC IV.
Attention based deep learning over free text discharge summaries to predict thirty day readmissions.
Compared ResNet50, VGG16, and InceptionV3 for CT based lung lesion classification on the IQ OTH NCCD dataset.
What our 2025 fellows said about their experience.
Tensor Lab showed me meaningful challenges that combine both clinical and technical domains. As I prepare for my PhD applications, this experience was invaluable in clearly defining my research direction.
Research Fellow, 2025Tensor Lab taught me to conceptualize and develop applications for LLMs that require significant domain knowledge.
Research Fellow, 2025The guidance from the faculty and mentors was exceptional. I wouldn't have reached our final publication goals without their help.
Research Fellow, 2025Collaborating with a global cohort provided me with a network of contacts for future projects.
Research Fellow, 2025