COMPLEMENTARY CLINICAL AND EXPLAINABLE ARTIFICIAL INTELLIGENCE MODELS FOR RISK STRATIFICATION OF CANCER-ASSOCIATED VENOUS THROMBOEMBOLISM

@article{abreu2026complementary,
  title = {{COMPLEMENTARY CLINICAL AND EXPLAINABLE ARTIFICIAL INTELLIGENCE MODELS FOR RISK STRATIFICATION OF CANCER-ASSOCIATED VENOUS THROMBOEMBOLISM}},
  author = {Márcia Fayad Marcodes ABREU and Mohamad Al BANNOUD and Sandra Martins SILVA and Stephany Cares HUBER and Beatriz MARTINELLI and Maria Carmen G FERNANDES and Júlio C TEIXEIRA and José Barreto C. TEIXIRA and Carmem Silva Passos LIMA and Nelson A ANDREOLLO and Maurício ETCHEBEHERE and Cláudio Saddy Rodrigues COY and Jose Claudio Teixeira SEABRA and Ricardo K MUSSI and Helder TEDESCHI and Tiago D MARTINS and Gabriel Y OTTAIANO and Rubens Maciel FILHO and Silmara A MONTALVÃO and Joyce Maria ANNICHINO-BIZZACCHI},
  journal = {Hematology, Transfusion and Cell Therapy},
  year = {2026},
  month = oct,
  doi = {10.1016/j.htct.2026.108078},
  url = {https://doi.org/10.1016/j.htct.2026.108078},
  note = {CC BY},
}

TY  - JOUR
TI  - COMPLEMENTARY CLINICAL AND EXPLAINABLE ARTIFICIAL INTELLIGENCE MODELS FOR RISK STRATIFICATION OF CANCER-ASSOCIATED VENOUS THROMBOEMBOLISM
AU  - Márcia Fayad Marcodes ABREU
AU  - Mohamad Al BANNOUD
AU  - Sandra Martins SILVA
AU  - Stephany Cares HUBER
AU  - Beatriz MARTINELLI
AU  - Maria Carmen G FERNANDES
AU  - Júlio C TEIXEIRA
AU  - José Barreto C. TEIXIRA
AU  - Carmem Silva Passos LIMA
AU  - Nelson A ANDREOLLO
AU  - Maurício ETCHEBEHERE
AU  - Cláudio Saddy Rodrigues COY
AU  - Jose Claudio Teixeira SEABRA
AU  - Ricardo K MUSSI
AU  - Helder TEDESCHI
AU  - Tiago D MARTINS
AU  - Gabriel Y OTTAIANO
AU  - Rubens Maciel FILHO
AU  - Silmara A MONTALVÃO
AU  - Joyce Maria ANNICHINO-BIZZACCHI
PY  - 2026
DA  - 2026/10/23
JO  - Hematology, Transfusion and Cell Therapy
DO  - 10.1016/j.htct.2026.108078
UR  - https://doi.org/10.1016/j.htct.2026.108078
AB  - Introduction: Cancer-associated venous thromboembolism (CAT) causes major morbidity and mortality, and current risk models show only moderate performance with limited biological interpretability. Combining clinical prediction with explainable artificial intelligence (AI) may offer complementary strategies for personalized CAT risk assessment. Objectives: To develop and internally validate two complementary approaches for CAT prediction: a pragmatic clinical risk score adaptable to D-dimer availability, and an explainable machine learning model integrating clinical and laboratory biomarkers. Materials and methods: A prospective multicenter cohort of 806 newly diagnosed cancer patients without prior chemotherapy was followed for six months; 47 CAT events occurred. Patients were split into derivation/training (70%) and validation (30%) cohorts. A weighted clinical score was derived by multivariable logistic regression, scoring: tumor site (pancreatic/gastric +3; lung, gynecologic, lymphoma, bladder, or testicular +2), adenocarcinoma/carcinoma histology (+3), platinum-based or gemcitabine chemotherapy (+1), diabetes mellitus or long-term insulin (+1), prior VTE (+1), platelets ≥150 × 10⁹/L (+1), leukocytes ≥9 × 10⁹/L (+2), hemoglobin ≤8 g/dL or erythropoiesis-stimulating agent use (+2), age ≥75y (+2), and D-dimer ≥1.44 mg/L when available (+2). Thresholds: without D-dimer, low risk <9, high risk ≥9; with D-dimer, low risk <10, high risk ≥10. Only D-dimer added independent predictive value among biomarkers tested. ML algorithms were compared via 10-fold cross-validation; LightGBM performed best. Performance was assessed by C-index, sensitivity, specificity, predictive values, and bootstrap CIs; SHAP was used for interpretability. Results: CAT patients more often had adenocarcinoma (59.6% vs 37.8%, p = 0.005) and pancreatic tumors (14.9% vs 3.2%, p < 0.001), higher D-dimer (median 1963 vs 699 ng/mL, p < 0.001), lower hemoglobin, and higher platelets/leukocytes. The clinical score reached validation C-indices of 0.65 without D-dimer and 0.68 with D-dimer (n = 469, 31 events), with specificity rising from 0.91 to 0.94 and positive predictive value from 0.23 to 0.33; negative predictive value stayed high (0.96-0.97). CAT incidence was 22-33% in high-risk versus 3-4% in low-risk patients. The explainable LightGBM model reached a validation C-index of 0.68, specificity of 0.82, and negative predictive value of 0.96. SHAP ranked D-dimer as the strongest predictor, followed by tumor histology, platelet count, hemoglobin, inflammatory biomarkers, diabetes mellitus, and age. Discussion: The two approaches are complementary: the clinical score is simple and robust even without D-dimer, while adding D-dimer improves classification where available. The explainable AI model gives individualized, transparent, biologically plausible predictions with differently weighted clinical and laboratory contributions. Both models reliably identified low-risk patients. Conclusion: These complementary clinical and explainable machine learning models achieved similar discrimination while addressing different clinical needs, offering practical, biologically interpretable strategies for CAT risk stratification, though external validation is still required before routine clinical use. Funding: Fundação de Amparo à Pesquisa do Estado de São Paulo (FAPESP, grant #2016/14172-6). Approved by the Research Ethics Committee, University of Campinas (CAAE 87212318.7.0000.5404).
SN  - 2531-1379
SN  - 2531-1387
ER  -