🩺 GPT sensitivity 88% for favorable ICH recovery
🩺 GPT sensitivity 88% for favorable ICH recovery
In a single-center cohort study of 409 patients with intracerebral hemorrhage, GPT-4o mini used routine admission data from the first 24 hours to predict 3-month functional outcome with 88% sensitivity for favorable recovery and an AUC of 0.86—comparable to two board-certified neurologists (AUC 0.85). The study found moderate correlation with actual outcomes (r=0.64), but GPT tended to underestimate 3-month disability, especially in older patients and those with milder deficits.
Why It Matters To Your Practice
Prognostication after ICH is often variable and vulnerable to bias, particularly early in admission when families and teams want guidance.
A commercially available LLM performed similarly to clinician averages using only routine first-day clinical data.
High sensitivity for favorable outcome suggests potential value as a decision-support tool when screening for patients who may recover better than expected.
Clinical Implications
GPT-4o mini predicted favorable 3-month outcome (mRS 0-3) with 88% sensitivity and 67% specificity.
Overall discrimination was comparable to neurologist predictions: AUC 0.86 for GPT vs. 0.85 for clinician-averaged scores.
GPT predictions correlated moderately with actual 3-month mRS (r=0.64), versus r=0.68 for clinician averages.
The model systematically underestimated disability (mean difference -0.99), which could affect counseling if outputs are taken at face value.
Insights
Performance dropped in older patients (r=0.58) and in patients with milder deficits (r=0.22), highlighting subgroup limitations.
One clinician underestimated outcomes while the other overestimated them, underscoring that human prognostication also varies.
The findings support feasibility, not replacement: AI may help standardize early estimates rather than serve as a standalone prognostic authority.
The Bottom Line
For ICH prognostication, GPT-4o mini was roughly on par with neurologists for identifying favorable 3-month recovery, with strong sensitivity but only moderate specificity.
Use AI outputs as a structured adjunct to clinical judgment—especially cautiously in older patients, milder presentations, and any high-stakes goals-of-care discussion.
Reference: Assessing the feasibility of ChatGPT in predicting outcomes after intracerebral hemorrhage.