Episode Summary
Executive Summary: Devin Singh described how his work as a pediatric emergency physician and clinical AI lead at SickKids evolved into founding Hero AI to translate machine learning into real hospital workflows. He emphasized end-to-end system design, bias mitigation, stakeholder collaboration, and practical automation that improves care, efficiency, and equity rather than producing models that only perform well on paper.
Main Topics: From pediatric ER medicine to clinical AI: Singh traced his path from medical school and residency to pediatric emergency medicine, explaining how difficult clinical encounters motivated him to use AI to improve outcomes across multiple layers of care delivery. Clinical workflow automation in the emergency department: He described using triage data and machine learning to trigger pre-emptive tests, such as x-rays, so work can be completed while patients wait, compressing delays and improving throughput. Translating research into real-world deployment: A major theme was the gap between publication-centric research and actual implementation, including the political, operational, legal, and institutional work required to move AI into practice. Hero AI and building deployable healthcare infrastructure: Singh explained that Hero AI exists to create the software, human-computer interaction, and technical pipelines needed to bring AI-driven automation into the hands of patients and providers. Bias, equity, and fairness in AI: He argued that AI can both reveal and reduce institutional bias, but only if models are audited across demographics and deployed responsibly with attention to ethnicity, age, gender, and other factors. Technical strategy: simple models first, advanced methods when needed: Singh noted that logistic regression often works well, but some problems require neural nets, explainability tools like SHAP, or data augmentation techniques such as GANs for rare conditions. Healthcare regulation, consent, and liability: He discussed a CIFAR workshop focused on legal and regulatory questions around consent, privacy, accountability, and how to govern constantly learning AI systems in healthcare.
Key Arguments: AI should be used to improve multiple points in care delivery at once—health literacy, triage, diagnostics, automation, and workflow efficiency—rather than solving a single isolated problem. The hardest part of clinical AI is not model performance but deployment: workflow integration, stakeholder alignment, and institutional politics. A two-pronged strategy can make automation safe and useful: let highly confident model predictions trigger action while misses fall back to standard human workflow. Publication-driven academic incentives often discourage the messy, practical work needed to translate models into actual patient benefit. Bias assessments are mandatory before deployment because models trained on historical clinical data can reproduce institutional inequities. AI can help standardize care and reduce disparities, for example by ensuring pain reassessment happens consistently across demographic groups. Simple models should be tried first; neural networks are not automatically the best choice, and explainability matters for clinical adoption. In low-data settings, dimensionality reduction and synthetic data generation can improve performance, but building multi-hospital data networks is a better long-term solution. Responsible healthcare AI requires early involvement from clinicians, engineers, ethicists, lawyers, and operational stakeholders. AI regulation must address special issues such as pediatric consent, privacy, liability, and continuous model updating.
Data Points: Years since AI realization: about 4 years ago - Singh said he first seriously encountered AI concepts after a conversation at a wedding roughly four years prior to the interview. Model capture strategy: top 20% of cases - For triage automation, he described only acting on the most confident subset of predictions while the rest follow normal workflow. Missed cases tolerated in automation design: up to 80% missed by the model - He said the system can intentionally miss many cases as long as it reliably captures a smaller high-confidence group that adds efficiency. False positive rate: 0.09% - Singh cited some models with extremely low false positive rates, though he noted even small errors can matter at hospital scale. Age window for craniosynostosis intervention: before 3 to 4 months of age - He explained that early detection of craniosynostosis matters because surgery is less invasive before this window closes. Data size for rare-disease imaging project: hundreds to a couple thousand images - He described a limited 3D imaging dataset collected over years for craniosynostosis prediction.
Pivotal Quotes: "AI. And at the time, I didn't even. This is like four years ago. I was like, what do you do? What is AI? I had no idea." — Devin Singh: He recalled first being pointed toward AI as a way to address the complexity of clinical improvement. "I will become the computer scientist, I will learn the language." — Devin Singh: He explained his decision to learn enough technical depth to bridge medicine and machine learning collaboration. "It would be totally irresponsible, in my opinion, to deploy machine learning solutions that influence care without bias assessments for age, gender, ethnicity." — Devin Singh: He was emphasizing the ethical necessity of auditing models for demographic performance differences before clinical deployment.
Implications: Clinical AI will matter most when it is embedded into hospital workflow, audited for bias, and designed with clinicians, lawyers, and operators from the start. The future is less about flashy models and more about safe, equitable automation that improves care at scale.