基于时间验证的机器学习模型预测非创伤性脊髓损伤步行能力
本研究回顾性纳入148例非创伤性脊髓损伤(NTSCI)患者,用入院早期临床指标构建九种机器学习模型并以GBDT优选,内部测试AUC为0.91(95%CI:0.77-1.00),时间外部验证37例准确率0.81(95%CI:0.68-0.92);作者强调结果仅为探索性,需多中心大样本确认后方可考虑临床应用。
要点
- GBDT模型内部测试AUC 0.91、召回率0.92、准确率0.90(均附95%CI)。
- 时间外部验证37例,准确率0.81、召回率0.70(均附95%CI)。
- 关键预测因子包括下肢总肌力、年龄和感觉评分等。
核心结论
- 该初步模型对NTSCI患者康复后步行能力概率(分类阈值0.5)具有潜在预测价值,但结果属探索性,临床应用前需更大规模多中心队列确认。
研究设计与样本
- 回顾性研究,含模型开发与时间外部验证;纳入148例NTSCI患者,随机分为训练集(80%)与内部测试集(20%)。
- 构建四种特征集并比较九种算法,依据AUC、召回率和F2分数选择梯度提升决策树(GBDT)并调参。
- 最终模型在37例NTSCI患者(2025年7月至2026年5月)队列上独立验证;获伦理批准(M202616)并豁免知情同意,回顾性注册(MR-11-26-021748)。
主要发现
- GBDT模型表现优于其他算法;内部测试AUC 0.91(95%CI:0.77-1.00)、召回率0.92(95%CI:0.71-1.00)、准确率0.90(95%CI:0.77-1.00)。
- 关键预测因子包括下肢总肌力、年龄等。
- 时间外部验证准确率0.81(95%CI:0.68-0.92)、召回率0.70(95%CI:0.40-1.00)。
关键数值与推荐
- 预测目标为康复后步行能力概率,分类阈值设为0.5。
- 模型优选依据AUC、召回率和F2分数,最终采用GBDT。
临床怎么用
- 早期入院临床指标结合机器学习或有助于识别最可能从强化康复中获益的NTSCI患者,支持个体化康复规划与资源分配。
- 当前模型应视为有前景的研究工具而非可直接使用的临床辅助工具。
局限性
- 模型在单中心开发,患者数量相对较少。
- 作者指出准确性需在更大规模多中心研究中确认后方可用于临床。
- 研究为回顾性设计,结果属探索性。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
BACKGROUND: Walking ability is the priority for spinal cord injury (SCI) patients. Nontraumatic SCI (NTSCI) is increasing with aging populations, yet its prognosis remains understudied. Early predictionF using admission clinical indicators could facilitate decision-making and rehabilitation planning. OBJECTIVE: To preliminarily develop and validate a machine learning model using early admission clinical indicators for predicting post-rehabilitation walking ability probability (with a classification threshold of 0.5) in patients with NTSCI. METHODS: This retrospective study included model development and temporal external validation. We enrolled 148 NTSCI patients, randomly allocating them into training (80%) and internal test (20%) sets. Four feature sets were constructed from objective clinical measures. Nine algorithms were compared; Gradient Boosting Decision Tree (GBDT) was selected based on AUC, recall, and F2-score, followed by hyperparameter tuning. The final model was validated independently on a cohort of 37 NTSCI patients (July 2025 to May 2026). Ethical approval (M202616) and consent waiver obtained; retrospectively registered (MR-11-26-021748); no patient involvement. RESULTS: The GBDT-based model outperformed all others. On internal testing, it achieved an AUC of 0.91 (95% CI: 0.77-1.00), recall 0.92 (95% CI: 0.71-1.00), and accuracy 0.90 (95% CI: 0.77-1.00). Key predictors included total lower limb muscle strength, age, and others. On temporal external validation, it showed encouraging performance with accuracy 0.81 (95% CI: 0.68-0.92) and recall 0.70 (95% CI: 0.40-1.00). CONCLUSION: This preliminary model may have potential for predicting walking ability in patients with NTSCI; however, these findings are exploratory and require confirmation in larger, multicenter cohorts before any clinical application can be considered. Can Artificial Intelligence Help Predict Walking Recovery After Non-Traumatic Spinal Cord Injury? A Machine Learning Model with Promising Early ResultsSpinal cord injury can severely affect a person’s ability to walk. Non-traumatic spinal cord injury—caused by tumors or blood vessel problems—is becoming more common as the population ages, yet doctors lack reliable tools to predict whether a patient will regain walking ability after rehabilitation. This motivated us to explore whether machine learning could help. We hypothesized that early neurological examinations—such as leg muscle strength and sensation—would strongly predict walking recovery. We analyzed data from 148 patients and developed a prediction model using machine learning, a method well-suited for capturing complex patterns among multiple clinical factors. We then tested the final model on a separate group of 37 patients. The model performed well: it correctly identified most patients who regained walking ability. The strongest predictors were leg muscle strength, age, and sensation scores—findings consistent with our clinical expectations. These results suggest that early neurological indicators, combined with machine learning, could help clinicians identify patients most likely to benefit from intensive rehabilitation, supporting more personalized care and better use of resources. However, the model was developed at a single center with a relatively small number of patients. Its accuracy needs confirmation in larger, multicenter studies before clinical use. At this stage, it should be viewed as a promising research tool rather than a ready-to-use clinical aid. Nonetheless, it offers a first step toward individualized rehabilitation planning for patients with non-traumatic spinal cord injury.
引用
- Tong A, Guo J, Cai X, et al. A Machine Learning Model With Temporal Validation for Predicting Walking Ability in Nontraumatic Spinal Cord Injury. Neurorehabil Neural Repair. 2026. PMID: 42885901. doi:10.1177/15459683261493914
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。