Abstract: FR-PO0012
From Big Data to Bedside: Evaluating a Predictive Model for Hospital-Acquired Thrombotic Microangiopathy
Session Information
- Artificial Intelligence and Digital Health at the Bedside
November 07, 2025 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Artificial Intelligence, Digital Health, and Data Science
- 300 Artificial Intelligence, Digital Health, and Data Science
Authors
- Alfieri, Carlo, Universita degli Studi di Milano, Milan, Lombardy, Italy
- Maselli, Giulio Paolo, Fondazione IRCCS Ca' Granda Ospedale Maggiore Policlinico, Milan, Lombardy, Italy
- Ippolito, Laura, Fondazione IRCCS Ca' Granda Ospedale Maggiore Policlinico, Milan, Lombardy, Italy
- Cantaluppi, Vincenzo, Azienda Ospedaliero Universitaria Maggiore della Carita, Novara, Piemonte, Italy
- Roccetti, Massimiliano, AegingTech, Rome, Italy
- Angelino, Andrea, AegingTech, Rome, Italy
- Martelli, Francesca, AegingTech, Rome, Italy
- Di Natale, Elisa, AegingTech, Rome, Italy
- Castellano, Giuseppe, Universita degli Studi di Milano, Milan, Lombardy, Italy
Background
Thrombotic Microangiopathy (TMA) is a severe, rare condition with high mortality if not promptly recognized. Diagnosing TMA in hospitalized patients (HP) is challenging. We developed and assessed a laboratory-based algorithm to support early detection of suspected TMA cases.
Methods
We conducted a retrospective analysis on 1736 hospitalizations from Novara hospital. An automated algorithm generated alerts based on specific laboratory parameters at admission (sCr>1.2 mg/dL, platelet count <150000, Hb<13 g/dL (M), <12 g/dL (F)) or significant trends during hospitalization (AKI, substantial platelet, or Hb reduction within 48h). Each hospitalization was independently evaluated and classified by clinicians into 3 groups: High Probability (score 7–10), Possible (4–6), and No TMA (0–3) with the aim to compare algorithm-generated alerts to clinician-assigned labels.
Results
1394 hospitalizations with adequate laboratory data were analyzed. Based on clinical assessment, 60% (n=839) were classified as No TMA, 34% (n=473) as Possible and 6% (n=82) as High Probability. The algorithm generated alerts in 77% of cases, achieving high sensitivity (94%), identifying nearly all clinically significant cases (99% High Probability, 94% Possible). Specificity was limited (34%), with 66% of No TMA cases triggering alerts, often due to transient or alternative conditions identified at admission. Positive predictive value was 49%, with accuracy of 58%. False positives frequently resulted from benign oscillations and were common in emergency or intensive care settings. High Probability cases had higher mortality (44%) versus Possible (24%) and No TMA (15%) cases, highlighting clinical relevance.
Conclusion
Our results show that while the algorithm provides a robust early-warning system to support clinician vigilance for potentially life-threatening TMA conditions. It has excellent sensitivity, confirming its potential as an effective clinical surveillance tool to identify patients at risk for TMA. While specificity remains limited, leading to false positives and increased clinical workload, integrating additional clinical data is likely to enhance accuracy. Future iterations of the algorithm, incorporating broader clinical parameters, may further optimize its utility in clinical practice.