Crash Risk Predictors in Older Drivers: A Cross-Sectional Study Based on a Driving Simulator and Machine Learning Algorithms

Abstract

The ability to drive depends on the motor, visual, and cognitive functions, which are necessary to integrate information and respond appropriately to different situations that occur in traffic. The study aimed to evaluate older drivers in a driving simulator and identify motor, cognitive and visual variables that interfere with safe driving through a cluster analysis, and identify the main predictors of traffic crashes. We analyzed the data of older drivers (n = 100, mean age of 72.5 ± 5.7 years) recruited in a hospital in São Paulo, Brazil. The assessments were divided into three domains: motor, visual, and cognitive. The K-Means algorithm was used to identify clusters of individuals with similar characteristics that may be associated with the risk of a traffic crash. The Random Forest algorithm was used to predict road crash in older drivers and identify the predictors (main risk factors) related to the outcome (number of crashes). The analysis identified two clusters, one with 59 participants and another with 41 drivers. There were no differences in the mean of crashes (1.7 vs. 1.8) and infractions (2.6 vs. 2.0) by cluster. However, the drivers allocated in Cluster 1, when compared to Cluster 2, had higher age, driving time, and braking time (p < 0.05). The random forest performed well (r = 0.98, R2 = 0.81) in predicting road crash. Advanced age and the functional reach test were the factors representing the highest risk of road crash. There were no differences in the number of crashes and infractions per cluster. However, the Random Forest model performed well in predicting the number of crashes.

Quality of life and socio-demographic factors associated with nutritional risk in Brazilian community-dwelling individuals aged 80 and over: cluster analysis and ensemble methods

“…The comparison between RF and XBGoost model allowed us to verify the consistency among the variables which are most important to predict nutritional risk, and the performance of these ensemble methods. Several set models have been tested by both algorithms to improve predictive ability, and they have been frequently used with good results in the healthcare field (78)(79)(80).…”