Clinical and generative AI, built responsibly
AI is a means to an end. It earns its place when the output is repeatable, traceable to its source, and saves someone time, and when it hands off to a person the moment it should.
- Where
- iCAD (2023 to 2025); CareTogether (2026 to present)
- My role
- Chief Product Officer; co-founder and CPTO
- Built
- FDA-cleared AI; production LLM agents
- Headline result
- First FDA AI clearance in 4 years; 4,000+ sites
Outcomes
- iCAD's first FDA AI clearance in four years, with AI in use at 4,000+ provider sites.
- Deployment cut from about 43 days to 3 hours by moving to cloud delivery.
- A 20-year research and commercial agreement with Google and a distribution partnership with Microsoft.
- Production LLM agents at CareTogether with a second model reviewing every call.
iCAD: FDA-cleared AI at the point of care
- Regulated AI end to end. Led the AI portfolio, including deep learning and machine learning models, from research and model development through clinical validation and FDA clearance.
- Data with provenance. Grew the training-data pipeline from about 7M to 20M+ mammography images from 100+ clinical sites, with full provenance for FDA review.
- Clinical research. Designed the clinical studies for breast cancer risk prediction and for breast arterial calcification (BAC), a marker of cardiovascular risk.
- Workflow, not another window. Integrated AI into PACS, RIS, and the EHR through DICOM, HL7, and FHIR, carrying results into the patient record.
- Decision-ready AI. Moved from showing findings to triaging cases, so radiologists read the highest-priority exams first.
CareTogether: generative AI with guardrails
- Grounded answers. Agents retrieve each member's care plan, medications, and history before responding; the data, not the model, is the source of truth.
- Three safety layers. In-call rules route crisis statements to emergency help and alert the family within seconds; a second model reviews every call; a human review queue handles high-severity findings.
- Consent and access. Each senior's information is visible only to their care circle, with recorded consent and audited access.
- Interoperability. Bidirectional EHR integration through FHIR and HL7.
- Governed engineering. AI-assisted development where every change is reviewed, tested, risk-tiered, and traceable before release.
What I look for in any AI product
- Is the output repeatable, and can a clinician see where it came from?
- Does it save time, or does every answer need fact-checking?
- Does it catch urgency and hand off to a person?
- Was it evaluated before it scaled, and is it monitored after?
References and related writing
- FDA clears ProFound Detection 4.0 (Radiology Business, November 2024)
- Google Health and iCAD extend their AI partnership to 20 years (Radiology Business, August 2023)
- iCAD's new FDA clearance (December 2024)
- Data-Driven Breast Care (Radiology Today, March 2025), citing my comments as iCAD's Chief Product Officer
- Do foundation models increase AI adoption in healthcare?
- Foundation models in radiology: it's all about the economics
- Where AI earns its value in radiology: decision support