S2415
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
PISA-derived stratifications generally outperformed both traditional and the stratifications from the previous study [1]. Key predictive factors identified in the symbolic expressions (VM, KPS, clinical profile) align with previous findings of [1]. PISA's improved performance appears to stem from using multiple thresholds in KPS depending on other input features, rather than a single threshold (as in [1]). An example patient stratification, chosen due to its relatively low complexity and good performance, is highlighted in blue in Figure 1 and shown in Figure 2.
the labour-intensive process of creating clinically useful models, comparing it with a previous manually designed, domain-specific approach[1]. Material/Methods: The internal dataset comprised 1043 patients from a single-centre retrospective cohort (2001-2010) [1]. For external validation, 339 patients were selected from the Dutch Bone Metastases Study [2]. Input features included Karnofsky Performance Score (KPS), visceral metastases (VM), brain metastases, number and location of spinal metastases, number of bone metastases, and clinical profile.PISA predicts survival by using GP-GOMEA [3] to multi-objectively optimise symbolic expressions (features), yielding interpretable feature sets that balance survival prediction performance and feature complexity (defined as expression size). PISA is agnostic to the underlying survival modelling technique. We consider both shallow (depth <= 3) survival trees and Cox regression in PISA. Each model is automatically converted into a meaningful patient stratification visualised through Kaplan-Meier curves and flowcharts linking input features to survival groups. C-index performance is measured by bootstrapping the external cohort 1000 times. Results:
Conclusion: In two large radiotherapy studies on patients with bone metastases, PISA achieves better performance than the previous study[1] using the same input features and replicating key findings. Therefore, our Pipeline for Interpretable-by-design Survival Analysis (PISA) could markedly reduce manual efforts traditionally required to create survival models for clinical practice, while delivering interpretable models with good performance. References: [1] Bollen, L. et al., 2014. Prognostic factors associated with survival in patients with symptomatic spinal bone metastases: a retrospective cohort study of 1043 patients. Neuro-Oncology 16(7), pp. 991-998.[2] Steenland, E. et al., 1999. The effect of a single fraction compared to multiple fractions on painful bone metastases: a global analysis of the Dutch Bone Metastases Study. Radiotherapy and Oncology 52, pp. 101-109.[3] Virgolin, M., Alderliesten, T. & Bosman, P.
Traditional baseline models were generally outperformed by PISA models (Figure 1). Similarly,
Made with FlippingBook - Share PDF online