Blood in the Code: Missed AI Alerts and Hidden Risks in Automated Anemia Management in Diabetic Patients

James Whitaker¹, Emily Clarke², Oliver Bennett³, Sophie Turner⁴

Abstract

Background:
Artificial intelligence (AI)-driven decision support tools are increasingly integrated into electronic health records (EHRs) for chronic disease management. However, their reliability and sensitivity in detecting complex comorbid conditions, such as anemia in type 2 diabetes mellitus (T2DM), remain underexplored.

Methods:
A retrospective cohort study analyzed 3,112 adult patients with T2DM managed in primary care and outpatient clinics across the NHS in England between January 2021 and June 2023. AI-based alerts for anemia screening were reviewed and compared to manual laboratory data audits. Primary outcomes included sensitivity of AI detection, prevalence of missed cases, and impact on clinical outcomes (referrals, hemoglobin trends, and hospitalization).

Results:
Of 3,112 patients, 724 (23.2%) had anemia based on WHO criteria. The AI system successfully flagged 63.8% of cases (n = 462), while 36.2% (n = 262) were missed. Among missed cases, 44.7% had Hb <10 g/dL, and 18.3% required transfusion within 6 months. Hospitalization for anemia-related complications was significantly higher in the missed group (9.9% vs 3.2%, p < 0.001). Notably, iron studies and B12 levels were omitted in 68.7% of unflagged cases. Systematic review revealed that the algorithm failed to integrate multi-source lab inputs and excluded microcytic indicators in 29% of false negatives.

Conclusion:
AI-based alert systems show promise in managing chronic disease comorbidities but are currently insufficient in reliably detecting anemia in diabetic patients. This study highlights critical blind spots in algorithm logic and data integration. Until these are resolved, clinicians must maintain active vigilance and not rely solely on automated alerts for patient safety.