Establishment Of Fall Risk Prediction Models For The Elderly In Rural Areas

ObjectiveBased on epidemiological field data,this study aimed to develop fall risk prediction models?tools?for the elderly population in rural areas,in order to strengthen health workers in rural areas of China to assess and predict the risk of falls among the elderly and to lay a solid foundation for early detection of high-risk populations and prevention for falls.MethodsIn field work,the prospective study design was conducted in Pingluo County and Qingtongxia City of Ningxia Hui Autonomous Region.12 administrative villages in Pingluo County and 15 administrative villages in Qingtongxia City were selected to participate in the study.Baseline information?including general demographic information,quality of life,history of diseases,usage of medications,physical measurements,and other tests?of 758elderly aged 60 years or older who met the inclusion and exclusion criteria of this study was collected.Three follow-up surveys were performed at 3^th,6^th,and 12^thmonth after the baseline survey,to collect whether fall has occurred during the follow-up period and specific information related to the fall.Subsequently,the machine learning principles and random sampling method were used to divide the data into training data sets and testing data sets at a ratio of 60%:40%.Logistic regression model,Cox proportional hazard regression model,and random forest model were established on the training data sets for analyzing risk factors and extracting feature importance ranking.The performances of the three models were evaluated and compared on the testing data sets by accuracy,sensitivity,specificity,positive predictive value,negative predictive value,and area under the receiver operating characteristic curve?AUC?.Results1 In baseline,a total of 758 participants were enrolled,including 350 females?46.17%?and 408 males?53.83%?.The average age of the participants was 66.95±5.12 years old,and66.28±4.69 years old for females,and 67.53±5.40 years old for males.2 In the first follow-up survey?3th months after participants were enrolled?,1participants lost to follow-up,and 62 participants experienced at least 1 fall,of which 54participants experienced 1 fall?7.13%?,6 participants experienced 2 falls?0.79%?,2participants experienced?3 falls?0.26%?,and the incidence of falls in the first follow-up survey was 8.19%.3 In the second follow-up survey?4th to 6th months after participants were enrolled?,1participants lost to follow-up,and 57 participants experienced at least 1 fall,of which 49participants experienced 1 fall?6.48%?,6 participants experienced 2 falls?0.79%?,2participants experienced?3 falls?0.26%?,and the incidence of falls in the second follow-up survey was 7.53%.4 In the third follow-up survey?7th to 12th months after participants were enrolled?,20participants lost to follow-up,and 103 participants experienced at least 1 fall,of which 73participants experienced 1 fall?9.89%?,23 participants experienced 2 falls?3.12%?,7participants experienced?3 falls?0.95%?,and the incidence of falls in the third follow-up survey was 13.96%.5 During the follow-up periods?within 12 months?,738 participants were successfully followed-up,and 20 participants were lost to follow-up,and 176 participants experienced at least 1 fall,of which 110 participants experienced 1 fall?14.90%?,43 participants experienced 2 falls?5.83%?,14 participants experienced 3 falls?1.90%?,5 participants experienced 4 falls?0.68%?,2 participants experienced 5 falls?0.27%?,2 participants experienced?6 falls?0.27%?,and the annual incidence of falls among the elderly in rural Ningxia was 23.85%.6 All variables of the baseline survey were used as input variables of the models,and logistic regression model,Cox proportional hazards regression model,and random forest model were trained on the training data sets.The results of the Logistic regression model in the testing data sets:accuracy was 79.66%,sensitivity was 19.67%,specificity was 95.30%,positive predictive value was 52.17%,negative predictive value was 81.99%,and AUC was0.617?95%CI:0.537⁰.696?.The results of the Cox proportional hazard regression model in the testing data sets:accuracy was 70.96%,sensitivity was 5.13%,specificity was 93.78%,positive predictive value was 22.22%,negative predictive value was 74.04%,and AUC was0.632?95%CI:0.565⁰.706?.Before training the random forest model,the parameters mtry and ntree of the random forest model were optimized.The results of tuning the parameters showed that when the parameter ntree was set to 500 and the parameter mtry was set to 3,the model performance was optimal.In the testing data sets,the results of the random forest model were:accuracy was 78.98%,sensitivity was 3.39%,specificity was 97.88%,positive predictive value was 28.57%,negative predictive value was 80.21%,and AUC was0.618?95%CI:0.562⁰.675?.Conclusion1 In this study,the predictive performance of Logistic regression model,Cox proportional hazards regression model,and random forest model were relatively general.Further study should be done in the future to find a model with stronger prediction performance.2 This study compared different models and found that Cox proportional hazards regression model was superior to Logistic regression model and random forest model.Cox proportional hazards regression model can evaluate and predict the occurrence risk of falls in the elderly to a certain degree,which can provide a reference for health workers in rural areas to carry out assessment and prediction of fall risk,and lay a foundation for further falls prevention.