71% of Hospitals Now Use EHR-Integrated Predictive AI in 2024 as Vendor Models Lead
Hospital use of EHR-integrated predictive AI hit 71% in 2024, up from 66% in 2023. Growth centers on billing and risk scoring, with vendor models leading and oversight improving.
Hospital Adoption of EHR-Integrated Predictive AI Jumps to 71% in 2024
Hospitals are moving from pilots to production. In 2024, 71% of hospitals used predictive AI integrated into their EHR, up from 66% in 2023, according to new federal data. Most are leaning on models from their EHR vendor, with billing and patient risk identification leading the use cases.
What the data covers
The report from ASTP/ONC tracks how hospitals use, evaluate and govern predictive AI-defined as statistical and machine learning models that classify or produce risk scores (readmissions, early disease detection, no-shows, treatment recommendations). Findings draw on the AHA IT Supplement surveys from 2023 and 2024.
The 2024 survey ran April-September 2024 across 2,253 non-federal acute care hospitals (51% response). The 2023 survey ran March-August 2023 across 2,547 hospitals (58% response).
Who is adopting fastest
- Medium and large hospitals outpaced small hospitals.
- Non-critical access hospitals led critical access hospitals.
- System-affiliated and urban hospitals led independent and rural hospitals.
- EHR vendor effect: among hospitals on the market-leading EHR, 90% used predictive AI in 2024 vs 50% for other EHRs.
Where hospitals are using predictive AI
- Revenue cycle: simplifying or automating billing was a top growth area.
- Operations: scheduling and throughput optimization.
- Care management: identifying high-risk outpatients to guide follow-up.
Source matters. Use of predictive AI for billing was higher with third-party or self-developed models (73%) than with EHR vendor models (58%). Identification of high-risk outpatients also grew faster among hospitals using third-party or self-developed AI than those using EHR vendor tools.
How hospitals source models
- 80% used models from their EHR developer in 2024.
- 52% used third-party models.
- 50% built models internally.
Most hospitals mix sources. EHR vendor models offer speed and integration. Third-party and in-house models offer flexibility and niche capabilities.
Evaluation and governance are catching up
- 82% evaluated models for accuracy.
- 74% assessed for bias.
- 79% conducted post-implementation evaluation or monitoring.
- Governance typically spans multiple entities: 66% reported a dedicated committee or task force; 60% cited division/department leaders. IT staff were least cited as the primary evaluation body.
Why this matters
Predictive AI is now a standard EHR feature set, not a side project. The next advantage comes from choosing the right use cases, validating models against local data and building guardrails that clinicians trust.
Practical steps for hospital leaders
- Prioritize 2-3 high-yield workflows: billing edits, readmission risk, no-show prediction, ED triage, care gap closure.
- Benchmark model performance locally. Validate AUROC/calibration on your population before go-live.
- Design human-in-the-loop workflows. Define who acts on a score, within what time window, and with what documentation.
- Track benefit and harm. Monitor precision/recall by subgroup, alert burden, time to action, downstream cost and outcomes.
- Use a tiered sourcing strategy. Start with EHR-native models for speed; add third-party or in-house where you need specialty depth or better fit.
- Stand up a cross-functional AI committee. Include clinical leaders, quality/safety, equity, compliance, privacy, security, revenue cycle and IT.
- Publish a model registry. List purpose, data inputs, training data provenance, validation metrics, monitoring plan and retirement criteria.
- Align with external frameworks (e.g., risk management, bias testing, documentation) and bake them into procurement and change control.
- Invest in training for clinicians, data teams and operations staff so they understand what a score means-and what it doesn't.
Key questions to ask your EHR and AI vendors
- What training data was used? How similar is it to our patient mix and care setting?
- Do you provide site-level calibration and drift monitoring out of the box?
- What are the model's top features/inputs and known failure modes?
- What subgroup performance gaps exist, and how are bias and fairness monitored?
- What is the expected clinical or financial ROI, and how will we measure it jointly?
- What safeguards exist for explainability, override, audit logging and rollback?
What good governance looks like
- Policy: clear criteria for model approval, re-approval and retirement.
- Process: pre-implementation testing, staged rollout, post-implementation surveillance.
- People: accountable owners for each model; escalation paths for issues; regular reporting to quality and executive committees.
- Equity: routine subgroup analysis; mitigation plans where gaps are detected.
- Security and privacy: data minimization, PHI handling standards, vendor risk assessments and BAAs.
Helpful resources
Build team capability
If you are formalizing AI education for clinicians, quality, or data teams, consider curated training to accelerate adoption and governance maturity.
