Explainable feature selection combining particle swarm optimisation with adaptive LASSO for MRI radiogenomics: Predicting HPV status in oropharyngeal cancer
| Authors |
|
|---|---|
| Publication date | 02-2026 |
| Journal | Computer Methods and Programs in Biomedicine |
| Article number | 109204 |
| Volume | Issue number | 275 |
| Number of pages | 11 |
| Organisations |
|
| Abstract |
Background: Radiogenomic modelling faces a significant challenge due to the high-dimensional nature of quantitative radiomic features and limited sample sizes. Feature selection is therefore essential to eliminate irrelevant features and mitigate overfitting. Particle swarm optimisation (PSO) has shown promise for effectively navigating large feature spaces, yet its effectiveness in radiogenomics remains unexplored. This study investigates the value of PSO-based methods, both independently and in combination with other advanced techniques, for MRI-based prediction of human papillomavirus (HPV) status in oropharyngeal squamous cell carcinoma (OPSCC). Materials and methods: Baseline contrast-enhanced T1-weighted MR scans from two centres were analysed: 153 patients in an internal cohort (randomly split into 80 % for training and 20 % for testing) and 157 patients in an external validation cohort. Radiomic features were extracted from manually segmented tumours and multiple feature selection methods, including PSO and its ensembles, filter-based methods, wrapper-based approaches, and shrinkage techniques were evaluated. Performance was measured and compared using the area under the receiver operating characteristic curve (AUC). Results: PSO alone had a reasonable predictive power on the internal test set (AUC = 0.76, 95 % CI: 0.57–0.92, p = 0.092). When combined with adaptive LASSO using Shapley values, PSO’s performance improved (AUC = 0.81, 95 % CI: 0.61–0.94, p = 0.023). Recursive feature elimination (RFE) selected the most relevant features (AUC = 0.91, 95 % CI: 0.79–1.00, p < 0.001). Despite this, RFE failed to generalise well to the external cohort (AUC = 0.52, 95 % CI: 0.42–0.60, p = 1). Meanwhile, the PSO–adaptive LASSO combination maintained a robust AUC = 0.78 (95 % CI: 0.70–0.85, p < 0.001), indicating superior generalisability. Conclusions: The explainable PSO–adaptive LASSO feature selection method provides generalisable radiogenomic signatures associated with HPV status in OPSCC, outperforming other feature selection approaches. This combination may serve as a robust strategy for developing transferable models in radiogenomics. |
| Document type | Article |
| Note | With supplementary material. |
| Language | English |
| Published at | https://doi.org/10.1016/j.cmpb.2025.109204 |
| Downloads |
1-s2.0-S0169260725006194-main
(Final published version)
|
| Supplementary materials | |
| Permalink to this page | |
