AI in Health: Saving Lives at Scale

Abstract

The session explored how artificial‑intelligence can be harnessed to diagnose silent heart attacks and other critical conditions at scale, especially for underserved populations. Ziad Obermeyer illustrated the promise of cheap handheld ECG devices combined with AI models trained on thousands of waveforms, and presented early field results from a Tamil Nadu health‑camp pilot that identified high‑risk individuals missed by traditional risk‑factor screens. A panel then examined practical pathways for scaling AI in health systems: Nura Health’s AI‑augmented caregiver‑support tool, the need for rapid yet rigorous evaluation designs, and governance frameworks that protect privacy while avoiding undue bottlenecks. The discussion concluded with actionable guidance on data quality, stakeholder engagement, and iterative product development.

Detailed Summary

1.1 Personal Story & Systemic Insight

• Anecdote: Obermeyer recounted a 2018 episode where, as a junior faculty member, he ignored worsening abdominal pain, ultimately discovering acute perforated appendicitis via CT scan at a local ER.
• Lesson: Even trained clinicians can miss life‑threatening conditions when they lack immediate diagnostic data; broader populations, especially those far from hospitals, face even greater information scarcity.

1.2 The Problem of Under‑Diagnosis

• Silent heart attacks: Cohort studies in the U.S. using cardiac MRI reveal that the majority of myocardial infarctions are “silent” – undetected by patients and clinicians.
• Global relevance: Similar under‑diagnosis exists for basic risk factors (hypertension, diabetes) in South India; limited access to cardiac MRI or ECG hampers detection.

1.3 Low‑Cost ECG Technology

• Device description: Handheld, $60 single‑lead ECG that connects via Bluetooth to a phone/tablet; requires no wires or specialized training.
• Constraint: Raw waveform alone is not actionable; traditionally requires physician interpretation, creating a bottleneck for remote use.

1.4 AI as the Missing Link

• Concept: AI models can map ECG waveforms directly to ground‑truth outcomes (e.g., prior silent MI) without physician adjudication.
• Data challenge: Obtaining labeled ground‑truth (e.g., cardiac‑ultrasound or MRI confirmation) is costly, especially in low‑resource settings.

1.5 Field Pilot in Tamil Nadu (Summer 2023)

• Design: Health camps adjacent to a district health centre; each participant received:
  1. Expensive reference tests – cardiac ultrasound (wall‑motion abnormalities) and, for some, cardiac MRI.
  2. Low‑cost tests – handheld ECG plus other sensors (pulse oximetry waveform, retinal photo, temperature, etc.).
• Data collection: Over 1,000 participants; each had paired high‑resolution and low‑cost measurements.

1.5.1 Early Algorithm Performance

• Risk stratification: Among the top 2–5 % highest‐risk scores, 10 % showed scar indicative of a prior silent MI on ultrasound, versus 0.2 % in the overall sample.
• Risk profile: Many flagged individuals lacked conventional risk factors (BP, cholesterol), highlighting AI’s ability to capture population‑specific patterns.

1.5.2 Cost‑Effectiveness

• Estimate: $2,000 per disability‑adjusted life‑year (DALY) averted, well within Indian cost‑effectiveness thresholds.

1.6 Live Demonstration

• Setup: Panelist Shreda Rajesh performed a live ECG on Obermeyer; waveform uploaded via web app; algorithm returned a low‑risk score within seconds.
• Take‑away: Demonstrated end‑to‑end workflow: finger‑placement → Bluetooth upload → AI inference → risk score.

1.7 Beyond ECGs – A Suite of Low‑Cost Sensors

• Pulse oximeter waveforms: Contain rich hemodynamic information.
• Aura ring (or similar wearables): Continuous temperature, photoplethysmography, activity data.
• Smartphone retinal imaging: $200 attachment yields CNS‑relevant images (stroke, intracranial hemorrhage).

1.8 Scaling Strategy & Future Directions

• Hardware package: <$20,000 total (ECG, retinal camera, portable ultrasound, chest X‑ray) could be deployed in a district kiosk.
• Data stewardship: Emphasized protecting Indian data “secret sauce” for safe, ethical, fast AI development.
• Evaluation pipeline: Immediate next step is a randomized controlled trial comparing AI‑enhanced ECG screening versus standard risk‑factor screening for detecting silent MI.

---

2. Panel Discussion – “Operationalizing AI for Health Systems”

2.1 Moderator Introduction (Aimee Barnes)

• Highlighted J‑PAL’s focus on rigorous impact evidence for AI solutions.
• Invited panelists to discuss implementation, evaluation, and governance.

2.2 Shahed Alam – AI‑Enhanced Caregiver Support (Nura Health)

Point	Detail
Problem addressed	Families often lack timely guidance after hospital discharge; health‑worker response times ballooned to >10 hours.
AI solution	Co‑developed an AI “co‑pilot” that surfaces relevant patient context and suggests evidence‑based replies to health‑worker queries.
Impact	Reduced average response time from >10 hours to ~30 minutes; increased caregiver retention (>30 % more repeat usage); health‑workers report higher job satisfaction.
Upcoming RCT	Will compare standard Nura Health caregiver program versus AI‑augmented version, focusing on newborn morbidity/mortality and caregiver engagement metrics.
Strategic vision	Use RCT results to advocate for national adoption of AI‑enhanced caregiver platforms as a new standard of care.

2.3 Zamir Bray – Evaluation Design for Rapid AI Innovation

Insight	Explanation
Speed vs. Rigor	Traditional RCT pipelines (6‑month design + 2‑year execution) risk being outpaced by AI model updates.
Adaptive/Pragmatic Trials	Propose short “safety‑signal” pilots (3–6 months) followed by adaptive designs that can incorporate model revisions without restarting the trial.
Policy Alignment	Evidence must be delivered in sync with policy windows; early, lower‑cost experiments can inform rapid decision‑making while larger trials run in parallel.
Human Factors	Adoption hinges on clinician trust; evidence should include qualitative assessments of user confidence and workflow integration.
Collaboration	Calls for joint effort among funders, local research institutions (e.g., ICMR), and organizations like J‑PAL to build evaluation ecosystems.

2.4 Rob Sherman – Governance, Privacy, and the “Human‑in‑the‑Loop” Myth

Issue	Recommendation
Governance as gate vs guide	Treat privacy and safety standards as design constraints, not blockers; embed privacy‑preserving techniques (differential privacy, anonymization) from the start.
Public willingness	Deliberative democracy studies in India & US show >90 % willingness to share health data for AI‑driven cures, provided safeguards exist.
Human‑in‑the‑loop re‑imagined	Recognize that AI can outperform humans on speed and consistency; humans should act as architects setting policy, reviewing edge cases, not as mandatory decision‑makers for every inference.
Building clinician trust	Demonstrations of AI catching critical errors (e.g., pointing out a fatal lumbar puncture mistake) can shift perception; iterative co‑design with clinicians is essential.
Regulatory foresight	Encourage transparent model documentation, auditable pipelines, and continuous bias monitoring to prevent systematic disparities.

2.5 Synthesis & Closing Remarks (Aimee Barnes)

• J‑PAL’s upcoming AI‑Health initiative (announcement slated for summit day 20) will fund research that blends rigorous impact evaluation with rapid AI prototyping.

• Key success factors identified by panel:

  1. Clear, patient‑centred problem definition (start with the need, not the tech).
  2. High‑quality, locally representative data (avoid “Western‑only” model bias).
  3. Iterative, adaptive evaluation designs that align with policy cycles.
  4. Governance built in, not bolted on – privacy‑by‑design and stakeholder trust.
  5. Human‑centered implementation – clinicians and caregivers must be co‑designers, not just recipients.

• Final audience takeaways: AI can unlock scalable diagnostics, but only when paired with robust evidence, ethical data practices, and thoughtful integration into frontline workflows.

Key Takeaways

• Under‑diagnosis is pervasive: Silent myocardial infarctions and many other acute conditions go unnoticed both in high‑income and low‑resource settings, creating a massive hidden disease burden.
• Low‑cost sensors + AI can bridge data gaps: Handheld ECGs (~$60) paired with AI models can flag high‑risk individuals who would be missed by conventional risk‑factor screening.
• Field pilot in Tamil Nadu shows promise: The AI‑enhanced ECG identified silent heart attacks at a rate 50× higher than the background prevalence, with cost‑effectiveness of roughly $2,000 per DALY averted.
• Randomized evaluation is essential: Both Obermeyer’s team and Nura Health are launching RCTs to rigorously compare AI‑augmented versus standard care pathways.
• Speedy, adaptive trials outperform classic RCTs for AI: Short “safety‑signal” studies and adaptive designs can deliver actionable evidence within policy windows.
• Data quality & locality matter: Models trained on Western data risk poor performance elsewhere; collecting representative, anonymized datasets in the target population is a competitive advantage.
• Governance should enable, not block: Privacy‑by‑design, transparent documentation, and public willingness to share data can coexist with rapid innovation.
• Human‑in‑the‑loop redefined: Clinicians should act as architects and overseers of AI systems, not as mandatory gatekeepers for each decision.
• Implementation hinges on people, not tech: Successful AI projects start with understanding caregiver and clinician workflows, then co‑design tools that fit naturally into existing practices.
• Iterative, small‑scale roll‑outs accelerate learning: Deploying pilots in limited geographies, monitoring outcomes, and iterating improves both algorithmic performance and user trust.

---

Prepared from the verbatim transcript of the “AI in Health: Saving Lives at Scale” session at the AI Summit, Delhi, 2026.

See Also:

• realizing-the-health-value-of-ai-safely
• harnessing-ai-for-equitable-and-resilient-health-systems
• empowering-billions-ai-health-records-and-the-future-of-people-powered-care
• ai-for-medical-imaging-and-clinical-diagnostics
• keynote
• harnessing-ai-for-health-equity-building-inclusive-human-capital-and-strengthening-researchindustry-collaboration
• enterprise-adoption-of-responsible-ai-challenges-frameworks-and-solutions
• ai-in-public-health-bridging-gaps-from-pilots-to-sustainable-impact
• diagnosing-the-future-ai-transforming-healthcare-systems
• ai-impact-forum-democratising-ai-resources