Skip to main content Skip to secondary navigation
Page Content

AIMI-HAI Partnership Grant

The AIMI-HAI Partnership Grant is designed to fund new and ambitious ideas that reimagine artificial intelligence in healthcare, using real clinical data sets, with near term clinical applications.

Current AIMI-HAI Partnership Grant projects:

Developing Artificial Physician Intelligence for the Early Detection of STEMI: Closing an Emergency Care Clinical Practice Gap

This project will advance the development of artificial intelligence (AI) to identify patients at risk for ST-segment elevation myocardial infarction (STEMI). Screening patients upon arrival in an emergency department is done to identify those who potentially have this most severe form of a heart attack. Diagnosing STEMI needs to occur within 10 minutes. The project team will improve current practice by pursuing the integration of AI, designed to replicate physician decision making, into STEMI screening. The execution of the project brings together 3 areas of expertise – emergency cardiovascular care, clinical informatics and predictive modeling analytics. The first phase of work will quantify the value of socio-demographic diversity characteristics in augmenting the sensitivity and precision of risk prediction. In addition, the team will silently pilot the screening model as physician AI within the Stanford Adult Hospital’s electronic health record (EHR) system. The AI will use live clinical care data for 6 months. The team will then measure the timeliness of the physician AI’s decision making, and the effectiveness of risk prediction in comparison to actual clinical care screening. This work explores the feasibility of a mechanistic approach to translating physician AI into the clinical environment in order to improve timely diagnosis for a time-sensitive medical condition.

Name Role School Department
Maame Yaa Yiadom PI School of Medicine Emergency Medicine
Ian Brown Co-PI School of Medicine Emergency Medicine

Digital Machine Learning Prediction Models for High-Value Oncology Diagnostic Testing

Comprehensive genome profiling of tumor specimens is an important new instrument in the diagnosis and treatment of cancer. With an increasing number of available molecular testing options, it can be difficult to choose the most relevant tests from the available test menu. Machine learning tools promise to be a new and important source of information for oncologists to make the best choice for their patients. For the Heme-STAMP tumor profiling test as an example application, we are developing a prediction tool that uses patient-specific data available in the electronic health record to predict how likely the molecular test will yield new and actionable results. This information will be presented in real-time to the ordering provider so that it can be used to select the most relevant test for the patient.

Name Role School Department
Henning Stehr PI School of Medicine Pathology
Dita Gratzinger Co-PI School of Medicine Pathology
Jonathan Chen Co-PI School of Medicine Biomedical Informatics

Opportunistic CT Imaging for Early Detection of Chronic Disorders: Multicenter Retrospective Validation & Prospective Deployment

Early detection of chronic disorders can improve population-level quality of life, longevity, and health care costs. While multiple screening tests for chronic disorders exist, these can have low compliance, add to healthcare costs, and be insensitive to early-stage diseases when interventions may be most effective. To address this, we will implement a solution for diagnosing ischemic heart disease, diabetes mellitus, and osteoporosis, using abdominal computed tomography (CT) scans that have already been acquired for additional reasons. Such CT scans can provide salient biomarkers such as the distribution of fat and muscle within the body, vascular calcifications, and bone mineral density measures, all of which are biomarkers of future disease activity. We will combine these images with a patient’s medical record and build explainable models for communicating model risks to both clinicians and patients. This high-value paradigm of opportunistic analysis using already-acquired imaging has the potential to improve patient outcomes without requiring additional testing or adding to the already burgeoning costs of healthcare.

Name Role School Department
Akshay Chaudhari PI School of Medicine Radiology
Marc Willis Co-PI School of Medicine Radiology
Daniel Rubin Co-PI School of Medicine Biomedical Data Science and Radiology
David Maron Co-PI School of Medicine Cardiovascular Medicine
Curtis Langlotz Co-PI School of Medicine Biomedical Data Science and Radiology
Robert Boutin Co-PI School of Medicine Radiology

Place Matters: The Streetscape Environment and Health among African Americans

Substantial literature demonstrates the significance of the human-made environment on key health behaviors and outcomes. However, most studies have been based on large-scale geographic (GIS) measures, which typically do not represent the local context in which individuals regularly interact with their environments. Evidence has emerged that the streetscape can affect health outcomes and disparities. Traditional streetscape audits require researchers to walk through an environment and manually classify features; however, this approach is time-consuming and relies on accurate and reliable human judgment. The emergence of widespread maps that feature panoramas of the environment (e.g., Google Street View) offers unprecedented opportunity for measuring streetscape features at the perspective from which individuals interact with their environment. Coupled with deep learning methods to extract features, this approach will overcome the limitations of the traditional streetscape audit. The overarching hypothesis of this work is that the presence of positive streetscape features can help enhance health. These types of features, such as lighting, safe pedestrian paths, and greenspace, may be especially important in under-resourced communities with high levels of health disparities. The proposed research will be conducted in collaboration with a population-derived cohort of African Americans living in the Deep South. Employing innovative human-centered artificial intelligence and computer vision methods, we will evaluate whether patterns of streetscape features are associated with physical activity, well-being, and chronic disease, independent of traditional risk factors and GIS-based measures.

Name Role School Department
Michelle Odden PI School of Medicine Epidemiology and Population Health
Abby King Co-PI School of Medicine Epidemiology
Sherri Rose Co-PI School of Medicine Primary Care and Outcomes Research
Jiajun Wu Co-PI School of Engineering Computer Science

Standardized Therapy Response Assessments of Pediatric Cancers

Imaging tests are essential for diagnosing cancers in children and for monitoring tumor response to therapy. New technologies enable simultaneous acquisition of positron emission tomography (PET) and magnetic resonance imaging (MRI) images, which allows for “one stop” cancer staging. However, the interpretation of 30,000 – 50,000 images generated with the PET/MRI technology is time consuming and prone to variability from one observer to another. In children with lymphoma, tumor response to chemotherapy is typically expressed by a 5-point score (the “Deauville score”) that describes the tumor signal on PET scans as being higher or lower compared to the signal of major blood vessels and the liver. Human observers tend to show limited reproducibility of intermediate scores of 2, 3 or 4. We propose to solve this problem by developing deep convolutional neural networks (Deep-CNN) that can accelerate and standardize pediatric PET/MR image data interpretation. The goal of our project is to develop a Deep-CNN for standardized Deauville scoring of lymphomas in children. We hypothesize that Deep-CNN can significantly (> 50%) speed up PET image interpretation times and improve the reproducibility of Deauville score assessments. To the best of our knowledge, this is the first attempt to apply Deep-CNN to interpretations of pediatric cancer imaging studies. Results will be readily translatable to the clinic and thereby, will have major and broad health care impact. Despite the obvious need of accelerated medical diagnoses for children with cancer, no current strategy has yet employed the use of Deep-CNNs to speed up and reduce variability in image data interpretation of children with cancer. This is because Deep-CNNs need to be trained on large data sets to achieve satisfactory performance. Since pediatric cancers are more rare than adult cancers and PET/MRI technologies are relatively new, there are limited PET/MRI data of children available to date. We are in a unique position to address this problem because we have established a centralized image registry with PET/MRI data of pediatric cancer patients from five major children’s hospitals. This will enable us to train and validate a Deep-CNN for therapy response assessments of pediatric cancers. Once established, our Deep-CNN can be made available to other institutions and cross-trained for other tumor types and adult patients.

Name Role School Department
Heike E. Daldrup PI Professor of Radiology (General Radiology) and, by courtesy, of Pediatrics Radiology
Daniel Rubin CO-PI Professor of Biomedical Data Science and of Radiology (Integrative Biomedical Imaging Informatics at Stanford), of Medicine (Biomedical Informatics Research) and, by courtesy, of Ophthalmology and of Computer Science Biomedical Data Science and Radiology

Self-service Data Science in the EHR with Multimodal Patient Embeddings

Analysing electronic health records (EHR) with machine learning holds great promise in tackling key problems in healthcare. However the scale, complexity, and heterogeneity of EHR data creates challenges for integrating these data into machine learning models. Current data science tools for EHRs largely focus on count-based models using structured data (e.g., medical codes, labs, demographics) and fail to capture critical information found in text and images. Moreover cohort sizes are typically small, failing to capture generalizable signals found across larger-scale patient populations. The inability to easily create feature representations that contextualize patient state and capture the full richness of EHR data directly impacts almost all clinical data science applications. Building on our prior work training foundation models using structured data from the entire Stanford Medicine patient population, we will develop a multimodal patient representation learning framework which combines structured EHR codes, clinical notes, and images. We will evaluate classifiers trained with our embeddings on three cohorts: pediatric sepsis patients presenting within 3 days of admission; patients diagnosed with pulmonary embolism, and CheXpert. This foundation model will be integrated into our patient search engine and cohort analysis tool, the Advanced Cohort Engine (ACE). We hypothesize that patient embeddings generated with multimodal data will improve classification performance across a range of clinical tasks, drive new insights via latent subclass analyses, and enable new modes of error analysis for clinical researchers. All of our code will be released as open source software and include guidance on using GCP infrastructure to train custom cohorts and include estimates for training time, cloud costs, and carbon footprint.

Name Role School Department
Keith Morse PI School of Medicine Pediatrics