Fig. 1: Workflow of the study.

A novel pathomics-based artificial intelligence (AI) platform was developed to assess cancer patient survival through a two-tiered predictive analytics approach, comprising patch-level deep learning (DL) and patient-level prediction. During patch-level DL, four convolutional neural networks were evaluated, and Visual Geometry Group19 (VGG19) was identified as the optimal algorithm according to the area under the curve (AUC) value. The features extracted by VGG19 were integrated into the whole slide image (WSI)-level pathomics signatures through Bag of Words (BoW) and Patch Likelihood Histogram (PLH) methods. These pathomics signatures were further selected by Correlation coefficients and Lasso-Cox regression to conduct the AI model. The Gradient-weighted Class Activation Mapping (Grad-CAM) was employed to highlight regions within the images that significantly influenced the model’s decision-making process. Two distinct DL algorithms were evaluated: weakly supervised DL (extensive labeling strategy, formed PathWS) and supervised DL (requiring annotations, formed PathS). PathS demonstrated better predictive efficacy according to Kaplan–Meier analysis in validation and external testing cohorts. Additionally, the clinical signature (CS) platform was developed based on clinicopathological parameters identified by Cox regression analysis. Furthermore, the integration of PathS with CS into a multimodal platform represented by a nomogram further enhanced predictive efficacy, suggesting strong potential for clinical application in oral squamous cell carcinoma management. AI artificial intelligence, WSI whole slide image, DenseNet Dense Convolutional Network, ResNet Residual Networks, VGG Visual Geometry Group, WSDL weakly supervised deep learning, SDL supervised deep learning, Grad-CAM Gradient-weighted Class Activation Mapping, BoW Bag of Words, PLH Patch Likelihood Histogram, PathWS pathomics-based AI platform developed through WSDL, PathS pathomics-based AI platform developed through SDL, c-index concordance index, CS clinical signatures, OS overall survival, AUC area under the curve, CI confidence interval.