AI Rends

Head and Neck cancers (HNC) are rare compared to breast, prostate and lung cancers. Prevalence is about 3-4% in European populations, and survival rate remains unacceptably poor globally (around 60% at 5 years in Netherlands, about 30% in India). HNC is highly heterogenous where treatment outcomes depend on etiology, genetics and carcinogen exposure e.g. tobacco and human papilloma virus (HPV). The most common subtype in developed countries is oropharyngeal squamous cell carcinoma (OPSCC). In spite of heterogeneity, treatment is seldom personalized. Routine care for locally advanced OPSCC is
neoadjuvant radiotherapy with or without chemotherapy, followed by clinical assessment 12 weeks after end of treatment. Major surgery i.e. Neck Dissection (ND) is a considered as an option to remove remnants of disease in the neck lymph nodes. However, patients' quality of life significantly degrades after ND due to speech impairment, dry mouth and eating difficulties. ND can be occupationally and psycho-socially debilitating.

There is a decades-long upsurge in OPSCC due to human papilloma virus (HPV) exposure in North America and Western Europe. Though cancer subtype prevalence varies geographically, there are concerns that tobacco consumption, urban intensification and HPV exposure will trigger similar epidemiological shifts in Eastern Europe, India and other rapidly-changing countries. Randomized studies with imaging follow-up suggest that invasive ND can be avoided in patients with complete response at 12 weeks. HPV-related OPSCC has a slower regression rate taking 4-5 months to complete, therefore ND might be offered after 12 weeks to patients who do not actually need it. However, ND has proven survival benefit in poor or partial (radiological) responders, but interpretation of images is strongly dependent on radiologists’ skill. Previous studies suggest ND as the safer option after equivocal clinical response because microscopic cancer remnants might evade radiological detection. In some patients, clinical equipoise exists because population outcomes demonstrate a trade-off between survival benefit and degraded quality of life. A knowledge gap now exists due to a lack of holistic clinical decision guidance that applies multiple factors to predict the utility of ND for an individual patient. Here is a strong clinical need where machine learning and artificial intelligence (AI) can be applied on massed real-world clinical datasets. We will develop and validate a holistic multifactorial clinical prediction model of OPSCC overall survival time and risk of cancer recurrence within 2 years. This will be used as clinical decision support to personalize a consultation whether or not to proceed with ND. We link retrospective clinical data and radiological imaging repositories in Netherlands, Poland and Canada to build a “distributed Deep-Learning network”. Pre- and post-treatment radiological images will be analysed with deep-learning neural networks (DNNs) and incorporated into holistic risk estimation models for mortality and recurrence.

Our data, models and software will be made Findable-Accessible-Interoperable-Reusable (FAIR) and disseminated in a scalable fashion, so that it can be easily re-deployed in future for new clinical questions. This clinical decision support model will open the way towards patient-focussed Shared Decision Making consultations, so each patient’s personal preferences and utility of invasive ND may be taken into account when choosing treatment.

This is a grant from Stichting Hanarth Fonds.