From prediction to practice: mitigating bias and data shift in machine-learning models for chemotherapy-induced organ dysfunction across unseen cancers
Contact us if these videos are not relevant.
Contact us if these videos are not relevant.
View abstract on PubMed
Summary
This summary is machine-generated.Machine learning models can predict chemotherapy-induced kidney and liver damage, optimizing monitoring schedules. These models generalize to new cancer types, improving patient care and treatment delivery.
Area Of Science
- Oncology
- Medical Informatics
- Machine Learning
Background
- Chemotherapy requires routine renal and hepatic function monitoring to prevent organ damage and treatment delays.
- Current monitoring frequency and timing are suboptimal.
- Machine learning (ML) deployment in clinical practice is hindered by model bias and data heterogeneity concerns.
Purpose Of The Study
- To develop ML models for individualized decisions on renal and hepatic monitoring timing during chemotherapy.
- To explore the impact of data shift on ML model performance.
Main Methods
- Utilized retrospective data from three UK hospitals.
- Developed and validated ML models to predict significant increases in creatinine and bilirubin levels post-chemotherapy cycle 3.
- Included 3614 patients undergoing treatment for breast, colorectal, lung, ovarian, and diffuse large B-cell lymphoma.
Main Results
- Achieved F2 scores of 0.7773 for bilirubin and 0.6893 for creatinine on unseen data, optimized for sensitivity.
- Demonstrated consistent performance on tumor types not included in training (F2 bilirubin: 0.7423, F2 creatinine: 0.6820).
- Highlighted ML's effectiveness in clinical settings and its generalizability to unseen tumor types.
Conclusions
- ML models show potential for improving chemotherapy patient care through optimized organ function monitoring.
- Proposed bias mitigation strategies including multi-site data evaluation, patient subgroup analysis, and formal bias measures.
- Data aggregation techniques were found to have unintended consequences on model bias.
Contact us if these videos are not relevant.
Contact us if these videos are not relevant.
