Comparing the performance of different machine learning methods to predict falls in acute care hospitals: a scoping review

File
Kenichi Goto, Kiyoshi Iwasa, Takuma Suzuki, Atsushi Okamoto, Yasuaki Murakami, HIsatsugu Taoka, Atsushi Okamoto
Purpose:

This study aimed to select and review studies comparing the predictive performance of falls in acute hospital patients using different machine learning techniques.



Methods:

Following the PRISMA guidelines for scoping reviews, systematic searches were conducted in databases including MEDLINE (PubMed), CENTRAL of the Cochrane Library, the largest Japanese search database (Ichusi), and AI search engines (Perplexity AI, Elicit, ResearchRabbit, Scite.Ai). The search focused on terms related to acute phase, falls, machine learning, and prediction.

Results:

We conducted preliminary searches and identified 530 papers, of which 9 were finally selected after title, abstract, and full-text screening and evaluation. The median (interquartile range: IQR) sample size across studies was 1100 (830-1500), with a median (IQR) fall rate of 11.3% (0.2-15). Studies used between 2 and 7 machine learning methods, including decision trees, baggings(random forest and bagging), boostings (XGBoost, LightGBM, and adaptive boosting), support vector machines, naive Bayes, neural networks, K-nearest neighbor, logistic regression and linear regression. The median (IQR) number of variables used in the models was 38 (18-69). To assess the prediction performance for falls, the area under the curve (AUC) was used in 8 out of 9 studies, with a median (IQR) AUC of 0.8 (0.7-0.9). The median (IQR) difference in AUC between machine learning methods within studies was 0.075 (0.045-0.085). When comparing the AUCs of the machine learning methods in each study, the ensemble methods based on decision tree algorithms tended to have slightly better predictive performance for falls.



Conclusion(s):

Our review of studies comparing the predictive performance of falls in acute hospitals using different machine learning methods showed overall good predictive ability across the methods used. However, there was no clear advantage among the machine learning models in terms of prediction accuracy.


Implications:

The results of this study highlight the potential of different machine learning approaches to predict falls in acute hospitals. Further research could focus on refining these models to improve predictive accuracy and clinical utility.



Funding acknowledgements:
No funds were used.
Keywords:
Machine learning methods
Falls In acute phase
Prediction
Primary topic:
Innovative technology: information management, big data and artificial intelligence
Second topic:
Neurology: stroke
Did this work require ethics approval?:
No
Has any of this material been/due to be published or presented at another national or international conference prior to the World Physiotherapy Congress 2025?:
No

Back to the listing