Episode Summary
Executive Summary: John Wong, a UCSF medical student and Gates Foundation AI researcher, discusses how healthcare data, incentives, and infrastructure shape the success of AI. He explains his Stanford work on EHR-based clinical decision support, the limitations of billing-optimized records, and why explainability and equity are critical. He also highlights global health AI challenges, especially data scarcity, deployment constraints, and the need to build local talent and community-centered solutions.
Main Topics: John Wong’s path into medicine, AI, and social impact: Wong traces his interest in medicine to personal and community experiences, then describes parallel growth in data science and deep learning, leading to work in healthcare disparities at UCSF and digital health at the Gates Foundation. How healthcare incentives shaped EHR design and clinician burden: The conversation outlines the U.S. healthcare system’s insurer-provider-patient incentive structure and argues that EHRs were built primarily for billing, contributing to clinician burnout and creating an opportunity for AI to reduce administrative work. Stanford deep learning project for clinical decision support: Wong describes a project using Stanford hospital data to recommend next-step orders for patients within 24 hours, aiming to assist clinicians by surfacing likely orders based on historical EHR patterns. Data engineering, feature selection, and practical model constraints: He explains that large EHR datasets required structured pipelines, heavy preprocessing, and feature reduction. Rather than feeding everything to a model, the team used frequency-based pruning and cost-aware experimentation to make training feasible. AI, bias, and healthcare disparities: Wong argues that retrospective EHR data can encode structural inequities, leading algorithms to reproduce disparities in resource allocation and performance across populations, especially marginalized groups. Global health and AI deployment in low-resource settings: At the Gates Foundation, he focuses on digital health infrastructure, offline-capable tools, and adaptation of AI systems to real-world constraints such as weak internet, poor lighting, and local workflows. Building local talent and community-centered AI: Wong emphasizes that sustainable health tech in LMICs requires local founders, local expertise, and long-term community engagement rather than imported solutions alone.
Key Arguments: Healthcare data is often optimized for billing rather than clinical utility, which makes it messy but also creates a major opportunity for AI to reduce administrative burden. AI in healthcare cannot be judged only on predictive accuracy; liability, explainability, and workflow fit determine whether it can be translated into practice. Retrospective EHR data reflects historical and institutional bias, so models trained on it can reproduce or worsen healthcare disparities. Feature reduction and structured preprocessing can outperform naive high-dimensional deep learning when compute is limited and signals are sparse. Healthcare AI must be evaluated across subgroups; average performance can hide serious failures for marginalized patients. In low-resource settings, deployment constraints such as internet connectivity, image quality, and electricity can be as important as model quality. Building local capacity and involving communities is essential for technologies to be useful and trusted in global health contexts.
Data Points: Stanford project dataset size: 55 million rows - Wong’s clinical decision support project used Stanford hospital data at very large scale. Patient count in project dataset: 100,000 patients - The Stanford EHR dataset covered roughly one hundred thousand patients. Physician burnout attributed to computer usage: 70% - Wong cites burnout among physicians as partly driven by EHR/computer work. Physician burnout prevalence: 50% - He says about half of physicians are burnt out today. Time physicians spend on computers: about half - Wong describes clinicians spending around half their work time entering data into computers. AI articles in PubMed: 100,000 results - Used to illustrate the breadth of general AI literature. AI + athlete articles in PubMed: 1,500 results - Serves as a contrast for more specific query volume. AI + vulnerable/refugee populations literature: less than 60 articles - Shows the scarcity of research on AI for vulnerable populations. Share of AI literature on vulnerable populations: 0.0006% - Wong characterizes this as the fraction of all AI articles represented by that niche. AI talent in nonprofit/healthcare: 4% - He cites a LinkedIn statistic suggesting very few AI professionals work in these sectors. Funding to local founders in East Africa: 5% - He says most tech funding does not go to local founders. Funding from outside the country in East Africa: 95% - The remainder goes to non-local founders or external actors. Rejected images in diabetic retinopathy pilot: 20% - In Thailand, poor image quality and lighting caused many scans to be rejected by the algorithm. Model output scale: over 1,000 clinical decisions - The Stanford recommendation system outputted a large set of possible orders/decisions. Rare recommendations frequency: 0.1% of visits - Items excluded from the output space were described as too rare to matter operationally. Projected vs allocated resources for Black Americans in one example: 18% allocated vs 46% projected - Wong cites a disparity example from risk-scoring/resource allocation systems.
Pivotal Quotes: "EHRs where we store all patient data optimized for billing, nothing else, which is ridiculous, right?" — John Wong: He explains why electronic health records are misaligned with clinical and AI needs. "There's a huge push in healthcare for explainable AI because of liability issues and things like that." — John Wong: He discusses why simple accuracy is insufficient for clinical adoption. "You have to engage with the community if you're ever gonna serve a community, I think." — John Wong: He argues that global health AI must be locally informed and community-centered.
Implications: Healthcare AI will succeed only if it is clinically useful, explainable, equitable, and deployable under real-world constraints. For industry and researchers, that means better data infrastructure, subgroup evaluation, and local partnerships—not just stronger models.