Tertiary lymphoid structures (TLS) have recently been shown to be predictive of survival in pancreatic adenocarcinoma (PDAC). This project aims to quantify and subtype TLS in three PDAC cohorts spanning over 600 patients. These findings will then be associated with clinical metadata, genomic mutations and transcriptional subtypes. The successful candidate will benchmark existing TLS identification methods and compare these to recently developed foundation models. Upon identification, we will attempt to stratify TLS into distinct subtypes based on the embeddings produced by foundation models. We will then attempt to identify whether these subtypes are driven by TLS specific aspects such as lymphocyte morphology or the surrounding environment such as the composition of the stroma or distance to the closest tumor. Finally, we will benchmark the extent to which these subtypes recapitulate transcriptional TLS subtypes we have already identified using spatial sequencing technologies. Upon creation of a robust TLS subtyping method, we will run it over slides from over 600 deeply phenotyped patients and associate the presence and TLS subtype with patient survival, genomic mutations and copy number aberrations as well as known transcriptional subtypes. Overall, this will be the most in-depth characterization of TLS’ in PDAC to date.

Predicting patient outcomes like risk of readmission is crucial for improving healthcare quality and efficiency. However, the unstructured, unstandardized nature of electronic medical record (EMR) data makes it challenging to develop robust supervised learning models. Large language models (LLMs) offer a promising approach to automating the curation of EMR data into machine-readable formats needed for predictive modeling. However, concerns remain around the reliability, stability, and tendency of LLMs to ”hallucinate” – generating plausible-sounding but factually incorrect outputs. In this project, the student will leverage open-source LLMs and explore prompt engineering and fine-tuning strategies to curate EMRs, mitigating the above issues and maximizing the effectiveness of LLMs for EMR data curation and predictive model development. The student will assess the performance of the LLM-powered approach against traditional manual data curation methods in terms of accuracy, scalability, and cost-effectiveness. The insights gained could enable more widespread adoption of LLM techniques to unlock the predictive power of EMR data at Sinai Health, leading to improved patient outcomes and healthcare system efficiency.