Machine learning to predict occult neck node metastasis in early oral squamous cell carcinoma - a comparison of models to risk stratify patients undergoing a sentinel lymph node biopsy.

BACKGROUND: Sentinel node biopsy (SNB) provides pathological staging of the neck in T1/T2 node-negative oral squamous cell carcinoma (OSCC). Up to 85% of SNB positive patients will have a negative completion neck dissection (CND), thus receiving unnecessary surgery. We sought to investigate the utility of machine learning to predict occult metastasis in sentinel lymph nodes (SLNs) and non-SLNs.

METHODS: A retrospective cohort study designed to collect clinical, pathological, and hematological markers was available from 177 cT1/2 cN0 OSCC patients who underwent SNB. Accuracy of logistic-regression (LR) and artificial-neural-network (ANN) machine learning models was compared to predict positive SLN and non-SLN metastasis (from CND) in SNB positive patients.

RESULTS: Overall, 38.4% (n = 68) were SNB positive. Of these patients, 17.6% (n = 12) had a positive non-SLN upon CND. LR and ANN models demonstrated 76.9% and 72.2% accuracies to predict SLN metastasis and 78.1% and 76.7% accuracies to predict non-SLN metastasis, respectively. Defining a low-risk sub-group for SLN metastasis with a probability of <0.3 increased the accuracy to 85%.

DISCUSSION: We provide evidence that AI machine learning can be harnessed within the SNB setting. While not meeting the predictive value required for immediate clinical use, there is strong potential to translate into future clinical pathways.