MACHINE LEARNING-AIDED EXPLORATION OF PHARMACOTHERAPY TO OVERCOME INAPPROPRIATE MULTIPLE MEDICATION AND IMPROVE MOTOR FUNCTION IN THE ELDERLY
*Keiichi Shigetome, Kentaro Oniki, Keiji Takata, Yuki Tateyama, Hiroki Yasuda, Miu Yokota, Sae Yamauchi, Norio Yasui-Furukori, Kazunori D Yamada, Junji SaruwatariAbstract
Background
Elderly individuals commonly suffer from several comorbid conditions and therefore use multiple medications. The loss of muscle mass and strength in the elderly is associated with increased risks of falls, fractures, and decreased activities of daily living. Several studies have shown that the use of multiple medications in elderly patients was associated with low muscle mass and physical performance. However, the impact of specific individual medications on motor function has not been fully evaluated, nor have effective procedures for reduction of medicine been established. Machine learning could be useful in this study, which analyze various patients’ factors such as age, gender, comorbidities, motor and cognitive function, and medications.
Aims & Objectives
To develop machine learning model that predicts motor function in elderly inpatients based on their clinical information and identify which medications that can positively or negatively affect their motor function.
Method
A retrospective observational study was conducted on 672 elderly patients (84.2 ± 7.9 years old) admitted to Sakurajyuji Hospital in Kumamoto, Japan from January 2017 to August 2019. Functional independence measure (FIM) scores, which consists of 13 motor items (mFIM) and 5 cognitive items (cFIM), at admission and discharge were obtained. mFIM score ranges from 13 to 91, with higher scores indicating better motor function. All patients were randomly divided into training (n = 470) and test (n = 202) sets. First, for the variable selection step, the relationship between discharge mFIM scores and each variable was examined in the training set using univariate analysis. Second, based on the selected factors, random forest (RF) model has been developed to predict discharge mFIM scores. Then, the SHaply Additive exPlanations (SHAP) method was employed to interpret the developed model.
Results
The mean of mFIM scores at admission and discharge were 45.5 ± 24.5 and 57.3 ± 26.5, respectively. As regards statistical analysis, 25 types of medications, including antihypertensives, hypnotics, antiepileptics, antidiabetics, laxatives, and analgesics were significantly associated with discharge mFIM scores in addition to age and history of cerebrovascular disease, Parkinson’ s disease, and pneumonia. The coefficient of determination (R2) and mean absolute percentage error of the RF model on the test set was 0.82 and 18.12%, respectively. According to SHAP analysis, admission mFIM and cFIM scores, and age strongly affected discharge mFIM scores. Additionally, non-steroidal anti-inflammatory drugs (NSAIDs) were positively associated with those scores, and oral corticosteroids were negatively associated with them.
Discussion & Conclusion
The developed RF model showed relatively good performance with R2 >0.8 on the test set. Oral corticosteroids could have negatively affected motor function in elderly patients. This may be related to protein catabolism with corticosteroids, which could induce fibrosis in muscle cells. Meanwhile, NSAIDs have been shown to could have a positive effect on motor function. Previous studies reported that the expression of prostaglandin F2α receptors, which are thought to be involved in muscle protein synthesis, increases with using NSAIDs. In conclusion, minimizing the use of oral corticosteroids may lead to improved motor function, and NSAIDs may support physical performance in the elderly.