IADR Abstract Archives
Deep Learning Approach for Predicting 5-Year Survival of Oral Squamous Cell Carcinoma Patients
Objectives: Oral Squamous Cell Carcinoma (OSCC) has an estimated prevalence of 1/10,000 to 10/10,000 and has been ranked by the World Health Organization as the eight most common malignant cancer in 2003. The aim of this pilot study was to create, train and validate a novel deep-learning based algorithm and utilize it to predict the 5-year overall survival rates of patients diagnosed with Oral Squamous Cell Carcinoma (OSCC).
Methods: The population-based Surveillance, Epidemiology, and End Results (SEER) registry was used to collect data on 10,196 patients diagnosed with primary OSCC between 2004-2010. 70% of cases were used to train our model and 30% was used to evaluate its performance. The predictive variables examined in the neural network were the primary site of cancer, extent of growth, age and sex of patient, grade and size of tumor at initial presentation. Statistical analyses were performed using IBM SPSS statistics version 25 (Armonk, NY: IBM Corp).
Results: Patients included in this study had an average age of 63.4 years (SD:14.0, Range 18-105 years). The average size of tumors was 23.64mm (SD:16.3mm) and 58% of subjects were male. In the final test, survival neural network model showed an area under the curve (AUC) of 0.733. The sensitivity of the model was 0.722 and the specificity was 0.614. The most important variables in order of importance in the algorithm were age (with a normalized importance of 100%), size, grade, extent of tumor, site and sex (with a normalized importance of 15.3%).
Conclusions: Given that the incidence of OPSCC has been projected to increase in coming years, and nearly 128,000 deaths worldwide are attributed to OSCC, there is great value in developing models to predict patient prognosis and survival outcomes. This exploratory study developed a pilot machine learning algorithm. Future work will be needed to provide patients with more effective and personalized prognostic evaluations.
Methods: The population-based Surveillance, Epidemiology, and End Results (SEER) registry was used to collect data on 10,196 patients diagnosed with primary OSCC between 2004-2010. 70% of cases were used to train our model and 30% was used to evaluate its performance. The predictive variables examined in the neural network were the primary site of cancer, extent of growth, age and sex of patient, grade and size of tumor at initial presentation. Statistical analyses were performed using IBM SPSS statistics version 25 (Armonk, NY: IBM Corp).
Results: Patients included in this study had an average age of 63.4 years (SD:14.0, Range 18-105 years). The average size of tumors was 23.64mm (SD:16.3mm) and 58% of subjects were male. In the final test, survival neural network model showed an area under the curve (AUC) of 0.733. The sensitivity of the model was 0.722 and the specificity was 0.614. The most important variables in order of importance in the algorithm were age (with a normalized importance of 100%), size, grade, extent of tumor, site and sex (with a normalized importance of 15.3%).
Conclusions: Given that the incidence of OPSCC has been projected to increase in coming years, and nearly 128,000 deaths worldwide are attributed to OSCC, there is great value in developing models to predict patient prognosis and survival outcomes. This exploratory study developed a pilot machine learning algorithm. Future work will be needed to provide patients with more effective and personalized prognostic evaluations.