Early Sepsis Prediction by Cascaded Classification of Multi-Modal Clinical Parameters

Tanuka Bhattacharjee1, Sakyajit Bhattacharya2, Varsha Sharma2, Anirban Dutta Choudhury2, Sunil Kumar Kopparapu2, Rupayan Chakraborty2, Upasana Tiwari2, Murali Poduval3, Sundeep Khandelwal3, Kayapanda Muthana Mandana4
1TCS Research and Innovation, 2TATA Consultancy Services Ltd., 3TATA Consultancy Services Ltd, 4Fortis Hospital


Physionet Challenge 2019 attempts to address the issue of early detection of sepsis, which is essential for providing timely and fruitful treatment to the patients. Our contribution lies in a cascaded prediction approach fusing both vital signs and laboratory values. There are thirty laboratory values, out of which seven or eight are the most significant factors for sepsis, as found in the medical literature. We thus divide the laboratory values in two subsets - subset A, consisting of the ones which have significant effects on sepsis, and subset B, with the rest. We extract different features including, number of times each value is measured, number of times a value goes out of its normal range etc., from each of the laboratory values from A and B. We use the set of features obtained from A to build a classifier for first stage of prediction. With the similar set of features from the subset B, we build the second classifier stage. If the classification on subset A does not give definitive prediction for a patient, (s)he is sent to the next level of classification, where the classifier on subset B is used. The fusion of these two stages gives a classification based on laboratory values. A similar path is formed with the vital signs employing features like mean, standard deviation etc. Finally the classification results obtained from laboratory values and vital signs are fused in an ensemble manner to come to a conclusion. Our algorithm for the unofficial phase of submission of the challenge yields an utility score of 0.13 on 5-fold cross-validation on the training dataset.