Tyndall National Institute, University College Cork, Lee Maltings Complex, Dyke Parad, T12R5CP, Cork, Ireland.
School of Nursing and Midwifery, University College Cork, Cork, T12 AK54, Ireland.
Aging Clin Exp Res. 2024 Sep 10;36(1):187. doi: 10.1007/s40520-024-02840-5.
The aim of this study is to explore the feasibility of using machine learning approaches to objectively differentiate the mobilization patterns, measured via accelerometer sensors, of patients pre- and post-intervention.
The intervention tested the implementation of a Frailty Care Bundle to improve mobilization, nutrition and cognition in older orthopedic patients. The study recruited 120 participants, a sub-group analysis was undertaken on 113 patients with accelerometer data (57 pre-intervention and 56 post-intervention), the median age was 78 years and the majority were female. Physical activity data from an ankle-worn accelerometer (StepWatch 4) was collected for each patient during their hospital stay. These data contained daily aggregated gait variables. Data preprocessing included the standardization of step counts and feature computation. Subsequently, a binary classification model was trained. A systematic hyperparameter optimization approach was applied, and feature selection was performed. Two classifier models, logistic regression and Random Forest, were investigated and Shapley values were used to explain model predictions.
The Random Forest classifier demonstrated an average balanced accuracy of 82.3% (± 1.7%) during training and 74.7% (± 8.2%) for the test set. In comparison, the logistic regression classifier achieved a training accuracy of 79.7% (± 1.9%) and a test accuracy of 77.6% (± 5.5%). The logistic regression model demonstrated less overfitting compared to the Random Forest model and better performance on the hold-out test set. Stride length was consistently chosen as a key feature in all iterations for both models, along with features related to stride velocity, gait speed, and Lyapunov exponent, indicating their significance in the classification.
The best performing classifier was able to distinguish between patients pre- and post-intervention with greater than 75% accuracy. The intervention showed a correlation with higher gait speed and reduced stride length. However, the question of whether these alterations are part of an adaptive process that leads to improved outcomes over time remains.
本研究旨在探索使用机器学习方法客观区分患者干预前后通过加速度计传感器测量的活动模式的可行性。
干预措施测试了实施虚弱护理套餐以改善老年骨科患者的活动、营养和认知能力。该研究招募了 120 名参与者，对 113 名具有加速度计数据的患者（57 名干预前和 56 名干预后）进行了亚组分析，中位年龄为 78 岁，大多数为女性。每位患者在住院期间都佩戴踝部加速度计（StepWatch 4）采集身体活动数据。这些数据包含每日汇总的步态变量。数据预处理包括步计数的标准化和特征计算。随后，训练了一个二进制分类模型。应用了系统超参数优化方法，并进行了特征选择。研究了两种分类器模型，逻辑回归和随机森林，并使用 Shapley 值解释模型预测。
随机森林分类器在训练期间的平均平衡准确率为 82.3%（±1.7%），测试集的准确率为 74.7%（±8.2%）。相比之下，逻辑回归分类器在训练时的准确率为 79.7%（±1.9%），测试时的准确率为 77.6%（±5.5%）。逻辑回归模型的表现优于随机森林模型，在验证集上的表现更好，表明其具有较小的过拟合。在所有迭代中，步长始终被选为两种模型的关键特征，同时还选择了与步速、步态速度和 Lyapunov 指数相关的特征，表明它们在分类中的重要性。
表现最佳的分类器能够以大于 75%的准确率区分干预前后的患者。干预措施与更高的步速和更短的步长相关。然而，这些变化是否是随着时间的推移导致结果改善的自适应过程的一部分，仍然是一个问题。
