Classification of Opioid Usage Through Semi-Supervised Learning for Total Joint Replacement Patients.

Opioid misuse and overdose have become a public health hazard and caused drug addiction and death in the United States due to rapid increase in prescribed and non-prescribed opioid usage. The misuse and overdose are highly related to opioid over-prescription for chronic and acute pain treatment, where a one-size-fits-all prescription plan is often adopted but can lead to substantial leftovers for patients who only consume a few. To reduce over-prescription and opioid overdose, each patient's opioid usage pattern should be taken into account. As opioids are often prescribed for patients after total joint replacement surgeries, this study introduces a machine learning model to predict each patient's opioid usage level in the first 2 weeks after discharge. Specifically, the electronic health records, patient prescription history, and consumption survey data are collected to investigate the level of short-term opioid usage after joint replacement surgeries. However, there are a considerable number of answers missing in the surveys, which degrades data quality. To overcome this difficulty, a semi-supervised learning model that assigns pseudo labels via Bayesian regression is proposed. Using this model, the missing survey answers of opioids amount taken by the patients are predicted first. Then, based on the prediction, pseudo labels are assigned to those patients to improve classification performance. Extensive experiments indicate that such a semi-supervised learning model has shown a better performance in the resulting patients classification. It is expected that by using such a model the providers can adjust the amount of prescribed opioids to meet each patient's actual need, which can benefit the management of opioid prescription and pain intervention.