Patient Similarity Using Population Statistics and Multiple Kernel Learning
; Proceedings of the 2nd Machine Learning for Healthcare Conference, PMLR 68:191-203, 2017.
We present a multiple kernel learning framework to learn similarity functions that compare physiological state between patients. A powerful ensemble kernel is learned from many base kernels evaluated on individual features. Our proposed framework captures two aspects of patient similarity: that patient similarity should be dependent on clinical context and that similarity should be modulated by the frequency and specificity of individual feature values. We validate our model on ICU data to predict hemodynamic instability and present analyses on using the patient similarity function to construct personalized cohorts. Our experiments show that the statistical properties learned by the kernels functions based on feature population distributions are significantly more predictive than naive stationary kernels (e.g. RBFs). Population-based kernels outperform RBF’s in identifying patient cohorts based on abnormality of their vitals and lab measurements and at predicting mortality.