Highlighted Projects

1. Predicting patient decompensation from continuous physiologic monitoring in the emergency department.

Published in NPJ Digital Medicine. [link]. Advised by David Kim MD PhD and Pranav Rajpurkar PhD. At the time of joining the bootcamp, Sameer Sundrani and Julie Chen were both undergraduates at Stanford. Image: Julie Chen presenting at SAEM 23.

Sameer Sundrani and Julie Chen first authored a research study on a framework called VitalML, which utilizes advanced machine learning techniques to predict the development of critical conditions in initially stable emergency department patients. By integrating traditional triage data with features derived from continuous physiologic monitoring, their models achieved exceptional accuracy in anticipating vital sign abnormalities like tachycardia, hypotension, and hypoxia within a 90-minute timeframe. These predictions surpassed the performance of models relying solely on standard triage data. This innovative approach not only enhances triage processes but also enables real-time monitoring of patients at risk of clinical deterioration. Sameer and Julie's research has the potential to revolutionize emergency and critical care, ultimately improving patient outcomes in a significant way.

2. Improving Radiology Report Generation Systems by Removing Hallucinated References to Non-existent Priors

Published in the Machine Learning for Health 2022 Conference [link]. Advised by Pranav Rajpurkar PhD. At the time of joining the bootcamp, Vignav Ramesh was first year undergraduate at Harvard and Nathan Chi was a first year undergraduate at Stanford. Image: Vignav Ramesh presenting, with Pranav Rajpurkar, at SAIL 23.

Vignav Ramesh and Nathan Chi first authored a project tackling the issue of hallucinated references to non-existent prior reports in deep learning models. First, they utilize a powerful language model called GPT-3 to rewrite medical reports without references to prior reports. This ensures that the generated reports are free from hallucinated references. Second, they employ a specialized language model called BioBERT for token classification, directly removing words that refer to prior reports. By applying these methods, Vignav and Nathan modify the MIMIC-CXR dataset, a collection of chest X-rays and their associated radiology reports, and retrain the CXR-RePaiR system. The resulting model, called CXR-ReDonE, outperforms previous methods in terms of clinical metrics. It achieves higher accuracy and produces radiology reports with significantly fewer hallucinatory references to prior reports. Their work marks a crucial step forward in the evolution of radiology report generation systems, bringing us closer to a future where advanced technology seamlessly supports medical professionals in providing optimal care to patients.

3. Multimodal Image-Text Matching Improves Retrieval-based Chest X-Ray Report Generation

Published in the Medical Imaging with Deep Learning 2023 Conference [link]. At the time of joining the bootcamp, Jaehwan Jeong was a second-year undergraduate at Stanford, Katherine Tian was an undergraduate at Harvard, Sina Hartung was a masters student in Biomedical Informatics at Harvard, & Andrew Li MD was a Gastroenterology fellow at Stanford. Image: Jaehwan Jeong giving NEJM award presentation at SAIL 23.

Jaehwan Jeong, Katherine Tian, Sina Hartung, and Dr. Andrew Li authored a research paper proposing an innovative approach to radiology report generation called X-REM (Contrastive X-Ray REport Match). The goal is to improve the retrieval-based generation of radiology reports by enhancing the coherence and accuracy of the generated text. Their approach, X-REM utilizes a novel image-text matching score to measure the similarity between a chest X-ray image and a radiology report for report retrieval. By employing a language-image model, the fine-grained interaction between the image and text is effectively captured, surpassing the limitations of cosine similarity often used in similar methods. The performance of X-REM surpasses multiple prior radiology report generation modules, demonstrating superior natural language and clinical metrics. Their work highlights the potential for significant advancements in radiology report generation, bridging the gap between AI-generated reports and those produced by human experts. With further development and evaluation, their research can contribute to the improvement of radiology practices, ultimately enhancing patient care in the medical field.

4. Adapt Segment Anything Models to Medical Imaging via Fine-Tuning without Domain Pretraining

Accepted for oral presentation at AAAI 2024 Spring Symposium for Clinical Foundation Models [link]. At the time of joining the bootcamp, Kevin Li was a third year undergraduate at Stanford. Image: Kevin Li giving an oral presentation at AAAI 2024 Spring Symposium series.

Kevin Li first authored a research paper examining the value of medical pre-training for foundation models in the context of parameter efficient fine-tuning. In particular, the goal is to examine whether an image segmentation foundation model with medical pre-training (MedSAM) necessarily outperforms the original model without medical pre-training (SAM) when both models undergo fine-tuning. The conclusion is that for certain fine-tuning approaches, the advantage of medical pre-training either is negligible or does not exist. An explanation for this behavior is the idea that the medical corpus MedSAM is pre-trained on is small in scale relative to the general corpus SAM is trained on, so that the advantage MedSAM derives in the zero-shot setting for medical benchmarks becomes negligible in the fine-tuning setting where task-specific supervision is paramount.