Radiomics-based causal machine learning for exploratory treatment-effect estimation of neoadjuvant chemotherapy cycle intensity in osteosarcoma: a proof-of-concept study.
Authors
Affiliations (5)
Affiliations (5)
- Department of Orthopedic Oncology, Changzheng Hospital, Second Military Medical University, Shanghai, 200003, China.
- Department of Radiology, Changzheng Hospital, Second Military Medical University, Shanghai, 200003, China.
- Department of Radiology, The 905th Hospital of the Chinese People's Liberation Army Navy, Shanghai, 200003, China.
- Department of Radiology, Changzheng Hospital, Second Military Medical University, Shanghai, 200003, China. [email protected].
- Department of Orthopedic Oncology, Changzheng Hospital, Second Military Medical University, Shanghai, 200003, China. [email protected].
Abstract
Radiomics-based modeling has shown promise for characterizing tumor heterogeneity, but its integration with causal machine learning for treatment-effect estimation remains underexplored in osteosarcoma. This study aimed to develop a proof-of-concept radiomics-based causal machine learning framework for exploratory estimation of average and individual treatment effects associated with neoadjuvant chemotherapy cycle intensity in osteosarcoma. This retrospective single-center study included 34 patients with osteosarcoma who underwent neoadjuvant chemotherapy followed by surgical resection. Radiomic features were extracted from pre-treatment T1-weighted magnetic resonance imaging and combined with baseline clinical variables. Three causal meta-learners-S-Learner, T-Learner, and X-Learner-were implemented to estimate counterfactual survival probabilities under high-cycle and low-cycle neoadjuvant chemotherapy strategies. Average treatment effects and individual treatment effects were derived from the predicted potential outcomes. The proposed framework enabled estimation of population-level and individualized treatment-effect measures using integrated radiomic and clinical covariates. The estimated average treatment effects differed in magnitude and direction across meta-learners, indicating instability of treatment-effect estimation in this small cohort. Confidence intervals crossed zero for two of the three learners, and model performance metrics were interpreted only as technical indicators of feasibility rather than evidence of generalizable predictive validity. This study demonstrates the methodological feasibility of combining radiomics with causal machine learning for exploratory treatment-effect estimation in osteosarcoma. Given the limited sample size, retrospective single-center design, treatment-group imbalance, missing outcome information, and uncertainty of causal assumptions, the findings should be regarded as hypothesis-generating rather than clinically actionable. Larger multicenter studies with standardized imaging protocols, adequate event counts, longer follow-up, and prospective validation are required before the translational relevance of radiomics-based causal treatment-effect estimation can be assessed.