From prediction to practice: mitigating bias and data shift in machine-learning models for chemotherapy-induced organ dysfunction across unseen cancers

• DOI: 10.1136/bmjonc-2024-000430
• Publication Date: 2024-11-03T03:06:54Z
• Authors: Matthew Watson; Pinkie Chambers; Luke Steventon; James Harmsworth ...; Angelo Ercia; Heather Shaw; Noura Al Moubayed
• Keywords: Clinical Practice

Objectives Routine monitoring of renal and hepatic function during chemotherapy ensures that treatment-related organ damage has not occurred clearance subsequent treatment is hindered; however, frequency timing are optimal. Model bias data heterogeneity concerns have hampered the ability machine learning (ML) to be deployed into clinical practice. This study aims develop models could support individualised decisions on while exploring effect shift model performance. Methods analysis We used retrospective from three UK hospitals validate ML predicting unacceptable rises in creatinine/bilirubin post cycle 3 for patients undergoing following cancers: breast, colorectal, lung, ovarian diffuse large B-cell lymphoma. Results extracted 3614 with no missing blood test across cycles 1–6 treatment. improved previous work by including predictions 3. Optimised sensitivity, we achieve F2 scores 0.7773 (bilirubin) 0.6893 (creatinine) unseen data. Performance consistent tumour types training (F2 bilirubin: 0.7423, creatinine: 0.6820). Conclusion Our technique highlights effectiveness settings, demonstrating potential improve delivery care. Notably, our can generalise types. propose gold-standard mitigation steps models: evaluation multisite data, thorough patient population analysis, both formalised measures performance comparisons subgroups. demonstrate aggregation techniques unintended consequences bias.