Sequence Based Prediction of Hospital Readmissions

We propose an N-gram based approach to make predictions about the next event in a sequence of hospital admissions. To predict the value of a variable in the next admission (e.g. length of stay), we use only the sequence of values of the same variable for all hospital admissions of the same patient so far (e.g. the length of all previous hospital stays) and no other information from the discharge record. We validate our method on inpatient data from the California Office of Statewide Health Planning and Development (OSHPD), for the prediction of 3 kinds of variables: whether the patient will be readmitted within 30 days, and what the associated length of stay and cost of the next admission will be. Our results show that in all cases 4-gram, 5-gram and 6-gram methods make more accurate predictions than the majority baseline method, by a margin that depends on the problem at hand as well as on the cohort, i.e. the patient population as a whole ("ALL-Cause") versus a congestive heart failure (CHF) cohort.