AI in Healthcare: 6 Key Updates You Need to Know

It’s December 24, 2025. Today brings major developments in AI-driven diagnostics, patient engagement, regulatory shifts, and clinical implementation. Below are six topics that matter most for physicians navigating AI in healthcare today:

Have suggestions? Reply to this email.

A 2025 review in Springer shows machine learning can detect acute kidney injury (AKI) risk before creatinine rises using routine EHR data. Common approaches include regression, tree-based models, boosting, and deep learning. The main gains are earlier warnings and better risk sorting compared to simple rules. Major limits include lack of external validation, calibration issues, data bias, explainability gaps, and workflow hurdles.

AKI is a core hospital safety and cost problem. Early prediction can shift care from reactive to proactive. KDIGO guidelines support this approach, but health systems need validated tools before scaling. Business cases require proven clinical benefit and documented cost or length-of-stay reduction.

Key actions: Demand externally validated, explainable models with real-world outcome proof before wide adoption.

Manatt reports 2025 as the year AI risks moved from theory to active legal and policy action. U.S. state legislatures and courts now act on AI governance, making oversight a business issue. This shift affects contracts, deal approvals, and compliance in healthcare transactions. Companies face legal exposure if AI use is not governed. Deals can be delayed or restructured for regulatory fit. State-by-state rules mean higher compliance cost and legal complexity.

Risk assessments and contract clauses are now operational needs, not planning exercises. Teams should treat AI governance as a deal risk and update procurement and M&A playbooks. Add clear AI warranties, audit rights, and liability caps in contracts. Build state-level compliance checks into transaction due diligence.

A 2025 structured review of 84 papers found many AI systems equal or beat experts in tasks like mammography, dermatology, and retinal screening. But a common finding shows models lose 15–20% accuracy when moved to other hospitals or populations. This gap links to narrow or biased training data.

AI can speed reads and cut costs, but gains do not transfer automatically. Deploying AI without diverse data and governance risks errors, liability, and inequity. Business leaders must validate externally and expect performance drops when moving sites. Require demographic and device data before purchase. Use the 6P lens—Performance, Provenance, Population, Privacy, Practice, Policy—for procurement and rollout.

A study in European Archives of Oto-Rhino-Laryngology tested Claude 3 Opus against ChatGPT 4.0 on 50 consecutive primary head and neck squamous cell carcinoma cases in March 2024. Claude 3 achieved superior diagnostic scores versus ChatGPT 4.0. For treatment planning, both produced similar recommendations.

AI choice can change diagnostic accuracy. Small changes in model performance affect triage, referral, and resource use. Early adopters should verify model performance on their own caseloads before clinical rollout. Test AI tools on local case mixes, use models as decision support only, and require documented validation and clinician oversight before deployment.

MedCity News reports health systems spend heavily on AI but struggle to engage patients. Amy Bucher of Lirio says true one-to-one ("N-of-1") personalization is needed. Health systems often use AI for broad, generic outreach that does not create real connections. Current AI spend is not translating into better patient action or outcomes.

Poor engagement lowers ROI on AI investments and can mean missed revenue, worse outcomes, and higher costs. Hospitals and vendors must prove AI drives real patient behavior, not just send more messages. Stop one-size-fits-all outreach and test truly personal messages. Tie AI tools to measurable engagement outcomes like appointments kept and medications taken. Start small pilots, measure results, then scale what works.

Mass General Brigham and Dana-Farber teams built an AI tool that reads CT scans to predict how many lymph nodes show extranodal extension (ENE). The model was tested on CT scans from 1,733 patients with oropharyngeal carcinoma. The AI score linked to higher risk of uncontrolled spread and worse survival. Adding the AI output to current clinical risk markers improved risk stratification.

This can help teams pick patients for treatment intensification or de-intensification. That can cut unnecessary toxicity and target resources to higher-risk patients. AI enables noninvasive ENE assessment from CT, avoiding upfront surgical staging. Integrate the AI with current risk models to sharpen patient selection. Evaluate local validation and regulatory details before clinical use.

This week shows AI is moving from promise to practice. Early warnings for kidney injury, better cancer diagnosis, and precise lymph node prediction are real. But the 15–20% performance drop across sites, governance turning legal, and patient engagement gaps remind us: validation, transparency, and true personalization matter as much as the algorithms themselves. Systems that act now on these fronts will lead. Those that wait risk liability, wasted spend, and missed clinical gains.