Umeå University's logo

Background: In Sweden around 1400 people are affected by head and neck cancer each year, and around 400 of these tumours are located in the mobile tongue (SCCOT). A major problem with these tumours is the high degree of relapse. In order to broaden our understanding of the group of squamous cell carcinoma of the head and neck (SCCHN) tumours we evaluated and compared the outcomes of panendoscopy with biopsy, ultrasonography with fine needle aspiration cytology (US-FNAC), and preoperative positron emission tomography/computed tomography (PET/CT) data from the same patients. As patients with SCCHN frequently have distant metastasis and locoregional recurrences, machine learning (ML) techniques were used to create classification models that accurately predict the likelihood of an early recurrence.

Materials and methods: From patients suspected of having head and neck cancer between 2014–2016 results from PET/CT, panendoscopy with biopsy and US-FNAC were compared. Clinical, genomic, transcriptomic, and proteomic markers identifying recurrence risk were investigated. In blood samples taken from healthy individuals, data from proteins relevant to inflammation and/or tumor processes were evaluated. The SHapley Additive Explanations (SHAP) approach was used to determine the best ML algorithm for feature selection. AdaBoost, Artificial neural networks (ANNs), Decision Tree (DT), eXtreme Gradient Boosting (XGBoost), and Support Vector Machine (SVM) were used to create prediction models. Clinical data from patients were analyzed using statistical and ML techniques.

Results: The concordance between results from PET/CT and panendoscopy with biopsy was 91.3%, and somewhat lower, 89.1%, for PET/CT and US-FNAC. The top contributors to classification with the ML approach were five mRNAs (PLAUR, DKK1, AXIN2, ANG, VEGFA), and 10 proteins (RAD50, 4E-BP1, MYH11, MAP2K1, BECN1, NF2, RAB25, ERRFI1, KDR, SERPINE1), using the extreme gradient boosting (XGBoost) method. The SHAP approach was used for feature selection. Using data from analysis of proteins in blood and interpretable ML showed that the Support Vector Machine (SVM) had the best performance with a balanced accuracy of 0.863, and a ROC-AUC of 0.924. The top three contributors to the SVM prediction model's performance were IL10, TNF Receptor Associated Factor 2 (TRAF2), and Kallikrein Related Peptidase 12 (KLK12). Recurrence was correlated with diabetes (p = 0.003), radiographic neck metastasis (p = 0.010), and T stage (p = 0.0012). A ML model got an accuracy rate of 71.2%. In the SCCOT group, diabetics predominated over non-diabetics, and also had lower recurrence rates and better survival (p = 0.012).

Conclusion: Results show that the combination PET/CT is useful in diagnosis of SCCHN. It further emphasizes the use of ML to identify transcriptomic and proteomic factors that are significant in predicting risk of recurrence in patients with SCCHN. It provides a methodical strategy for early diagnosis of SCCOT before onset of clinical symptoms using multidimensional plasma protein profiling and interpretable ML. A model for predicting recurrence of SCCOT is provided by ML utilizing clinical data. As SCCOT patients with co-existing diabetes showed a better prognosis than non-diabetics, results suggest that individuals with SCCOT, regardless of diabetes status, may benefit from therapeutic management of glucose levels.