USING MACHINE LEARNING TO DEVELOP MODELS FOR THE PREDICTION OF UPPER GASTROINTESTINAL CANCERS

ABSTRACT

Introduction Machine learning methods have been used to develop predictive models in gastroenterology.1 Previously we identified features including age, history of psychological disorders and severity of dysphagia symptoms which were correlated with upper gastrointestinal (UGI) cancers.2 We sought to create a machine learning based model which could be used to predict the presence of UGI in patients referred for endoscopy. Methods Patients were recruited as part of the Saliva to Predict rIsk of disease using Transcriptomics and epigenetics (SPIT) study. Patients were recruited from 2-week wait suspected UGI pathway referrals at 20 hospitals in the United Kingdom. We enriched the cohort with additional patients admitted with confirmed oesophageal adenocarcinoma. 60% of the data was used for model generation with 10-fold cross validation, while the models were tested on the remaining 40% of the data. We used seven methods to generate our models Linear Discriminant Analysis (lda), Classification and Regression Tree (cart), k-Nearest Neighbour (knn), Support Vector Machines (svm), Random Forest (rf), Logistic Regression (glm) and Regularised Logistic Regression (glmnet). Model performance was assessed using area under the receiver operating characteristic curve (AUC) and DeLong test was used for model comparison. Results 93 cancer and 715 non-cancer patients were included. The best three models with 18 features were glmnet, lda and glm which all achieved an AUC of greater than 0.80 (figure 1). For the testing dataset, AUC was 0.75 (95%CI 0.67- 0.83), 0.74 (95%CI 0.66-0.82) and 0.75 (95%CI 0.68-0.83) (p=ns for all 3 pairwise comparisons) respectively. When applying a cost function, the three models all achieved a sensitivity of 0.973 and a specificity of 0.234 to 0.388 for the testing dataset. Conclusions Our models compare favourably with the Edinburgh Dysphagia Scale, which has a sensitivity and specificity of 0.984 and 0.093 respectively.3 Our models have the advantage of an improved specificity, which could equate to fewer endoscopies being performed for low risk patients. Given rising waiting lists as a direct result of COVID-19, our tool could be used to prioritise patients who should be investigated sooner.4 We plan next to validate our models on a validation cohort to assess its generalisability.