Machine learning of patient similarity: A case study on predicting survival in cancer patient after locoregional chemotherapy

Identifying historical records of patients who are similar to the new patient could help to retrieve similar reference cases for predicting the clinical outcome of the new patient. Amongst different potential applications, this study illustrates use of patient similarity in predicting survival of patients suffering from hepatocellular carcinoma (HCC) treated with locoregional chemotherapy. This study used 14 similarity measures derived from relevant clinical and imaging parameters to classify the HCC patient pairs into two classes, namely the difference between their survival time being longer or no longer than 12 months. Furthermore, this paper proposes and presents a patient similarity algorithm for the classification, named SimSVM. With the 14 similarity measures as input, SimSVM outputs the predicted class and the degree of similarity or dissimilarity. A dataset was collected from 30 patients, forming 300 and 135 patient pairs as training and test datasets respectively. The trained SimSVM with linear kernel gave the best accuracy (66.7%), sensitivity (64.8%) and specificity (67.9%) on the test dataset.