MED-TMA

MED-TMA applies ensemble machine learning to provide probability-based differential diagnoses for thrombotic microangiopathy, aiming to improve diagnostic accuracy in settings with limited on-site ADAMTS13 testing and where the PLASMIC score performs variably across ethnic groups.

Key Features:

• Ensemble modeling: Integrates four selected machine learning methods chosen based on Area Under the Curve (AUC) performance and inter-method correlation.
• Feature selection: Employs a rigorous feature elimination process that identified five key clinical variables for prediction.
• Performance metrics: Reported AUCs of 0.945 on the development dataset and 0.924 on the validation dataset.
• Probability-based outputs: Produces continuous probability scores for differential diagnosis rather than only binary labels.
• Dataset and validation: Trained and validated on 319 primary TMA patients from 31 hospitals in Korea with a development set (D-set, n=212) and validation set (V-set, n=107).

Scientific Applications:

• Differential diagnosis of TMA: Supports differentiation among causes of thrombotic microangiopathy, including ADAMTS13-deficient thrombotic thrombocytopenic purpura (TTP).
• Clinical decision support without ADAMTS13 testing: Provides probabilistic diagnostic guidance when on-site ADAMTS13 testing is unavailable.
• Complementing PLASMIC score: Offers an alternative or complementary approach where ethnic variability reduces PLASMIC score performance.
• Risk stratification and triage: Enables prioritization of cases using continuous probability outputs to inform clinical management.

Methodology:

Development used data from 319 primary TMA patients across 31 hospitals in Korea (D-set n=212, V-set n=107); a feature elimination process selected five clinical variables; an ensemble of four ML methods was built based on AUC and inter-method correlation, yielding AUCs of 0.945 (D-set) and 0.924 (V-set).