Multimodal AI-based risk stratification for distant metastasis in nasopharyngeal carcinoma

The TNM (tumour-node-metastasis) staging system is the primary tool for treatment decisions in nasopharyngeal carcinoma (NPC). However, therapeutic outcomes vary considerably between patients, and guidelines for the management of distant metastasis treatment remain limited. This study aimed to develop and validate a deep learning-based risk score to predict NPC survival.

We developed a graph for nasopharyngeal carcinoma (GNPC) risk score, a multimodal deep-learning-based digital score incorporating signals from both haematoxylin and eosin-stained tissue slides and clinical information. Digitised images of NPC tissue slides were represented as graphs to capture spatial context and tumour heterogeneity. The proposed GNPC score was developed and validated on 1949 patients from two independent cohorts.

The GNPC score successfully stratified patients in both cohorts, achieving statistically significant results for distant metastasis (P < 0.001), overall survival (P < 0.01), and local recurrence (P < 0.05). Further downstream analyses of morphological characteristics, molecular features, and genomic profiles identified several factors associated with GNPC-score-based risk groups.

The proposed digital score demonstrates robust predictive performance for distant metastasis, overall survival, and local recurrence in NPC. These findings highlight its potential to assist with personalised treatment strategies and improve clinical management for NPC.