AI Healthcare 2026: Diagnosis, Drug Discovery, Hospital Ops

By April 2026 the FDA had authorized close to 1,500 AI/ML-enabled medical devices, and the agency cleared 27 of them in April alone — roughly one device per day. Five years ago the number was a few dozen. AI in healthcare has finished its proof-of-concept era. What’s happening in 2026 is integration, billing, and the unglamorous work of making the technology survive contact with hospital operations.

This piece walks through where AI for health actually sits in 2026 — radiology, pathology, drug discovery, ambient scribes, EHR integration, consumer cardiology, mental health, and the policy scaffolding underneath all of it.

Diagnostic radiology: where AI quietly ate the workflow#

About 77% of FDA-cleared AI/ML medical devices are in the radiology subspecialty. That’s where the imaging volume is, where the labeled training data is, and where the clinical workflow tolerates an assistive second reader.

The named vendors#

Aidoc — triage and prioritization across CT and X-ray. Multiple FDA clearances; integrated into PACS at hundreds of US hospitals.
Rad AI — radiologist reporting workflow (impressions, follow-up tracking, peer review). Pitched as productivity, not diagnosis.
Annalise.ai — chest X-ray and CT brain with broad finding coverage; growth in Australia, UK NHS pilots, and the Middle East.
Viz.ai — large vessel occlusion stroke detection. The reference case for AI in acute care, with documented reduction in door-to-treatment times when the pipeline is set up properly.

The thing to notice is that the production wins are workflow wins, not “AI replaced the radiologist” wins. Aidoc reorders the worklist. Viz.ai pages the on-call neurointerventionalist. Rad AI drafts the impression. Diagnostic accuracy is necessary but not sufficient; the value is in saved minutes per case across millions of cases per year.

Pathology catches up#

Pathology lagged radiology by years because digitizing slides is harder than ingesting DICOM. That gap closed. PathAI and Paige.AI both have FDA clearances and commercial deployments for prostate cancer detection. Pathology AI is now a credible second-reader use case, especially in regions with pathologist shortages.

AI in healthcare 2026 stack

Drug discovery: from speculation to Phase II readouts#

For a decade “AI drug discovery” meant computational screening with a press release attached. In 2025-2026 the field produced actual clinical readouts.

Insilico Medicine and rentosertib#

In June 2025 Insilico Medicine published Phase IIa results in Nature Medicine for rentosertib (ISM001-055), a TNIK inhibitor for idiopathic pulmonary fibrosis. The target was identified by generative AI; the compound was AI-designed. The 60 mg once-daily arm showed a mean +98.4 mL forced vital capacity change versus a -20.3 mL decline in placebo over 12 weeks. The trial enrolled 71 patients across 22 sites in China. This is the first proof-of-concept clinical validation of generative-AI-driven drug discovery to clear peer review at that level.

Recursion + Exscientia, now one company#

Recursion and Exscientia completed their merger on November 20, 2024 — Exscientia became a wholly owned subsidiary of Recursion. The combined company carries more than ten clinical and preclinical programs internally, more than ten partnered programs, and existing collaborations with Roche-Genentech, Sanofi, Bayer, and Merck KGaA. Recursion shareholders own about 74% of the combined entity.

AlphaFold 3 and Isomorphic Labs#

AlphaFold 3 shipped in May 2024, extending structure prediction to proteins, DNA, RNA, and ligands. Isomorphic Labs (the DeepMind spinout) raised $600M led by Thrive Capital in March 2025 and has active research collaborations with Eli Lilly, Novartis (expanded in February 2025), and Johnson & Johnson. Demis Hassabis said at the January 2026 World Economic Forum that Isomorphic’s first clinical trials are expected by the end of 2026. In February 2026 the lab released a technical report on IsoDDE, its drug design engine.

Tempus and precision medicine#

Tempus continues to operate the largest US clinical and molecular data platform underpinning AI-driven oncology decision support. Their value isn’t a single model — it’s the volume of multimodal patient data feeding the analytics layer.

Ambient scribes: the clearest mainstream AI win in clinical workflow#

If you ask hospital CIOs what AI product actually shipped to clinicians in 2025-2026, the answer is ambient documentation. The room records, the model drafts the note, the physician edits and signs.

The named players#

Abridge — deepest Epic integration; Highmark prior-auth partnership; broad health system rollouts.
Suki — multi-EHR (Epic, Cerner/Oracle, athenahealth, Elation), strong on multi-language and noise handling.
DeepScribe — independent ambient platform with multi-EHR support.
Notable — broader workflow automation including ambient documentation.
Epic’s own ambient scribe — Epic shipped a first-party offering, changing the build-vs-buy math for Epic-heavy systems.
Oracle Clinical Digital Assistant — Oracle’s voice + multimodal capture inside the Oracle Health EHR.

The ambient scribe market generated around $600M in 2025, with adoption reaching roughly 35% among large US health systems and more than 40% of US physicians using some form of AI documentation tool. The driver isn’t accuracy hype — it’s physician burnout, and CFOs noticing throughput.

The EHR integration question#

Epic, Oracle Health (Cerner), and athenahealth are the platforms that matter. Abridge is Epic-first with broader EHR support added via standards. Suki spreads more evenly. The integration pattern that ships is FHIR plus an EHR-native sidecar — not a separate app the clinician has to context-switch into.

Consumer and remote cardiology#

The Apple Watch + AliveCor stack pulled cardiology partway out of the clinic. ECG features on consumer wearables now feed clinician workflows for atrial fibrillation screening at population scale. Insurance reimbursement and integration into care pathways are the limiters, not the device itself.

The relevant 2026 trend is the consumer-device data feeding back into hospital systems as a structured input — not as a novelty.

Mental health AI: cautious mainstream#

The named platforms — Woebot, Wysa, Lyra, Spring Health — operate as employer benefits and clinically-supervised pathways rather than autonomous therapy. The post-2024 regulatory and clinical posture is to use AI for triage, between-session support, and clinician augmentation, not as a substitute for a licensed provider. The bar for safety claims rose sharply after several public incidents in consumer chat products in 2024-2025.

Hospital operations: the unglamorous high-value layer#

The AI work that pays the bills in hospital operations is rarely the work that makes the news.

Where the value is#

Bed management and length-of-stay prediction — discharge planning gets started earlier; capacity is freed.
OR scheduling and turnover optimization — small percentage gains in OR utilization compound to material revenue.
Revenue cycle: prior authorization and denials — generative AI drafts authorizations, surfaces evidence, appeals denials. Highmark and Abridge’s prior-auth partnership is one named example; the underlying pattern is widespread.
Supply chain and pharmacy demand forecasting — fewer stockouts, less waste.
Workforce scheduling — nursing ratios optimized against acuity.

These are the places where the operating margin actually moves. They’re also where the AI looks the most like “intelligent forecasting + workflow integration” and the least like a chatbot.

Hospital operations AI dashboards

Regulation: the FDA AI/ML Action Plan and PCCP#

The FDA’s AI/ML Action Plan and the Predetermined Change Control Plan (PCCP) framework are how the agency is trying to keep up with adaptive AI. PCCP lets a manufacturer pre-specify the kinds of model updates that won’t require a new submission, provided the update sits inside the agreed envelope. Adoption is gradual; the framework is the agency’s pragmatic answer to the fact that AI software does not behave like a static device.

Equity and bias#

Real cases — the historical underperformance of pulse oximetry on darker skin tones, dermatology models trained on overwhelmingly light-skinned datasets, algorithmic referral systems that under-prioritized Black patients — keep bias on the table as a clinical safety issue, not just an ethics talking point. The credible vendors publish subgroup performance; the credible buyers require it.

Casgevy: a one-and-a-half year update#

Vertex and CRISPR’s Casgevy (exagamglogene autotemcel) was FDA-approved in December 2023 for sickle cell disease. 2024 revenue was about $10M. 2025 full-year revenue was $115.8M, with about 111 new patients beginning infusions versus a much smaller cohort the prior year. In December 2025 Vertex presented data in children ages 5-11, with 100% of sickle cell patients achieving vaso-occlusive crisis-free status (mean duration 35.3 months). Rollout remains slower than the original projections because of cell-collection capacity and authorization workflows — not the science. The Casgevy story is a useful reminder that breakthrough therapies still depend on operational plumbing.

International angle#

UK NHS — ambient AI scribe pilots running across multiple trusts; the NHS AI Lab continues to coordinate procurement.
Singapore HealthHub — national digital health platform integrating AI-powered services on top of a unified citizen record.
Saudi Arabia MoH — large digital-health transformation under Vision 2030, with AI radiology procurement at national scale.
India ABDM — the Ayushman Bharat Digital Mission is building the identity and consent rails AI vendors need to operate at population scale.

The pattern: countries with a single-payer or coordinated procurement layer move AI adoption faster than the US, where each integrated delivery network buys independently.

Where pdpspectra fits#

Most hospital systems we work with do not need another AI vendor — they need the integration, governance, and operational scaffolding that turns vendor pilots into shipped clinical workflows. Our work concentrates on three layers.

Clinical data and EHR integration — FHIR pipelines, de-identification, vendor-neutral feature stores, model output back into the EHR (data engineering).
LLM and ambient scribe integration — Epic / Oracle / athenahealth interop, prompt and tool governance, audit logging (AI & LLM integration).
MLOps and monitoring — drift, bias, subgroup performance dashboards, PCCP-aligned change tracking (ML & MLOps).

For hospital operations work — bed management, OR scheduling, revenue cycle automation — we lean on business automation on top of those layers.

Banking AI Roadmap: What to Build First in 2026 — the same sequencing logic applied to a different regulated industry.
Open-Source LLMs in Production — the build-vs-buy frame that matters when you can’t send PHI to a hosted API.
Bedrock vs OpenAI vs Anthropic for Enterprise — model selection when BAAs and data residency are non-negotiable.

The 2026 picture isn’t that AI replaced clinicians. It’s that AI quietly moved into the worklist, the dictation, the prior auth, the bed board, and the discovery pipeline — and the buyers got serious about governance, equity, and integration. If you’re sizing an AI program inside a hospital, payer, or biotech and want a second pair of eyes on sequencing, tell us about the program.