Advanced survival prediction in head and neck cancer using hybrid machine learning systems and radiomics features

Abstract

Accurate prognostic stratification of Head-and-Neck-Squamous-Cell-Carcinoma (HNSCC) patients can be an important clinical reference when designing therapeutic strategies. We set to predict 4 outcomes: overall survival (OS), distant metastasis (DM), locoregional recurrence (LR), and progression-free survival (SP). We studied Hybrid Machine Learning Systems (HMLS), applied to datasets with radiomics features. In this multicenter study, 408 HNSCC patients were extracted from The Cancer Imaging Archive (TCIA) database. PET images were registered to CT, enhanced, and cropped. 215 radiomics features were extracted from each region of interest via our standardized SERA radiomics package. We employed multiple HMLSs: 12 feature extraction (FEA) or 9 feature selection algorithms (FSA) linked with 9 survival-prediction-algorithms (SPA) optimized by 5-fold cross-validation, applied to PET only, CT only and 4 PET-CT datasets generated by image-level fusion strategies. Datasets were normalized by z-score-technique, and cindices were reported to compare the models. For OS prediction, the highest c-index 0.73 ± 0.10 was obtained for HMLS with Ratio of low-pass pyramid (RP) fusion technique + gaussian process latent variable model (GPLVM) + causal structure learning-based feature modification method (CSFM). For DM prediction, we achieved 0.80±0.06 via Dual-tree complex wavelet transform (DTCWT) fusion + Laplacian Score (LAP) + Logistic regression hazards (LH). For LR prediction, we arrived at a c-index of 0.73 ± 0.13 using PET + Sammon Mapping Algorithm (SM)+ deep neural network to distribute first hitting times (DHS). For SP prediction, the performance of 0.68 ± 0.02 was obtained via PET + SM + Relative risk model-depend on time (CoxTime). When no dimensionality reduction (FEA/FSA) was employed, the above 4 performances decreased to 0.69 ± 0. 10, 0.74 ± 0.13, 0.66 ± 0.15, and 0.68 ± 0.04 for OS, DM, LR and SP prediction. We demonstrated that using fusion techniques followed by appropriate HMLSs, including FEAs/FSAs and SPAs, improved prediction performance.