Prediction models:
Machine Learning
vs Cox model

Background

This research project considers various predictive models for modeling osteosarcoma survival and disease progression. Various methods are examined, two of which are statistical approaches and extensions of the well-used Cox proportional hazard (PH) regression model, and two of which are data-based approaches, using machine learning techniques. Over the last decade interest in and publications on machine learning (ML) approaches in medical and specifically cancer research has grown, with analyses chiefly involving data with large numbers of covariates (large p), such as genomic data. A very limited amount of this has focused on survival outcomes specifically, and even less on clinical data with a small p. The limited research in this area coupled with the increasing popularity of ML methods in general has prompted a study into the potential of ML for analyzing clinical data with a small p.

Aims

Two applications – a dynamic prediction model and multi-state model - are performed with an aim towards clinical inference. The purpose is to obtain information not available from a simple Cox model, and evaluate this contribution of information. More concretely for the multi-state model and dynamic prediction model the respective objectives are as follows:

• Identifying prognostic factors for which the effect changes along with the progression of the disease. For example a predictor may be more or less predictive if a patient has experienced some intermediate event than when still in starting state.

• Quantifying the effect of intermediate events and time since surgery on 5-year survival predictions, and identifying prognostic factors whose effects vary over follow-up. 

An in-depth comparison of the two ML methods – artificial neural networks and random survival forests –is performed in an application and simulation study with the purpose of identifying advantages and pitfalls and the global objective of gaining insight into which methods may be suitable for clinical data with a small p. Of particular interest is the trade-off between flexibility and interpretability and model stability. Concrete aims are:

• Identifying/developing performance measures suitable for assessing individual model fit and comparing model fit across methods. 

• Using these performance measures to investigate the potential of ML for clinical data with a small p.

• Performing a simulation in a unique data based approach, with the purpose of gaining insight into performance measure behaviour and into the variable selection process. 

Relevance for cancer research

Within this project, research will be carried out in order to comprehensively establish the potential of ML for survival analysis of cancer data. Methodology developed will be applied to clinical trial data from the EURAMOS-1 study. The methodology of this research could be applicable to any kind of cancerous diseases.

Project Outcomes

• Spreafico M, Hazewinkel AD, van de Sande MAJ, Gelderblom H & Fiocco M (2024). Machine learning versus Cox models for predicting overall survival in patients with osteosarcoma: a retrospective analysis of the EURAMOS-1 clinical trial data. Cancers, 16(16), 2880. doi: 10.3390/cancers16162880

• Spreafico M, Hazewinkel AD, Gelderblom H & Fiocco M (2024). Dynamic prediction of overall survival for patients with osteosarcoma: a retrospective analysis of the EURAMOS-1 clinical trial data. Current Oncology, 31(7):3630-3642. doi: 10.3390/curroncol31070267

• Hazewinkel AD, Lancia C, Anninga J, van de Sande MAJ, Whelan J, Gelderblom H & Fiocco M (2022). Disease progression in osteosarcoma: a multistate model for the EURAMOS-1 (European and American Osteosarcoma Study) randomised clinical trial. BMJ Open, 12:e053083. doi: 10.1136/bmjopen-2021-053083

Team

• Prof. dr. M. Fiocco, Mathematical Institute Leiden University, Department of Biomedical Data Sciences Leiden University Medical Center & Princess Máxima Center for Pediatric Oncology 

• Prof. dr. H. Gelderblom, Department of Medical Oncology at the Leiden University Medical Centre 

• Dr. A.D. Hazewinkel, Department of Medical Statistics of the London School of Hygiene & Tropical Medicine (UK)

• Dr. M. Spreafico, Mathematical Institute Leiden University & Department of Biomedical Data Sciences Leiden University Medical Center