基于美国联邦医疗保险索赔数据的机器学习难以预测全关节置换术后30天非计划再入院的个体风险,但可揭示与年手术量相关的群体层面关联
研究分析2016年10月至2018年9月美国联邦医疗保险索赔数据中679,041例全关节置换术(TJA),使用广义相加模型预测30天非计划再入院。模型区分度欠佳(AUROC约0.65-0.66),不适于临床个体预测;但部分依赖图显示,既往365天外科医生和医院TJA手术量较低与再入院风险升高存在非线性群体关联,并存在手术量阈值。
要点
- 基于索赔数据的非线性机器学习模型对TJA术后30天再入院的个体预测区分度欠佳,AUROC仅0.65-0.66。
- 再入院风险随既往365天外科医生手术量增加而降低,约至75例TJA后不再进一步降低。
- 医院手术量增加至约600例TJA后,THA再入院风险不再进一步降低。
核心结论
- 大型联邦医疗保险索赔数据集与机器学习不足以提供临床可用的TKA或THA术后30天非计划再入院个体预测模型,提示需要索赔数据库未常规收集的其他因素来预测再入院。
- 识别出低外科医生和医院手术量与再入院风险升高之间的非线性群体关联,包括特定手术量阈值,超过该阈值后再入院风险不再降低,这可能间接有助于指导政策及患者选择医院或外科医生。
研究设计与样本
- 分析2016年10月至2018年9月全国联邦医疗保险索赔数据中接受住院TJA的患者,共679,041例TJA(239,391例THA,439,650例TKA)。
- 模型特征包括人口学、县水平健康社会决定因素、既往365天医院和外科医生TJA手术量、以及临床分类软件精炼的诊断和操作类别。
- 使用广义相加模型(含成对交互项)进行训练和评估,以AUROC和精确率-召回曲线下面积(AUPRC)在两个留出样本上评价预测性能。
- 所有入院(除最后2个月外)随机80%/20%拆分,训练队列为随机80%样本下采样后(含全部再入院和等量随机非再入院),随机20%样本作为第一测试队列(随机留出),最后2个月入院作为第二测试队列(2个月留出)。
主要发现
- 随机留出样本中,THA的AUROC和AUPRC分别为0.65和0.087,TKA分别为0.66和0.077;2个月留出样本中,THA分别为0.66和0.087,TKA分别为0.65和0.075。
- 最具预测性的特征(按平均绝对得分)及其部分依赖图显示与再入院风险升高的群体关联包括:患者年龄较大、既往365天外科医生和医院TJA手术量低、男性、有心臟诊断史且无肿瘤诊断、县水平可门诊护理敏感疾病住院率较高。
- 部分依赖图显示外科医生和医院手术量的非线性群体关联:THA和TKA再入院风险随外科医生既往365天手术量增加而降低,至约75例TJA(TKA的OR=1.2,THA的OR=1.3),但更高年手术量未见进一步风险降低。
- 对于THA,再入院风险随医院手术量增加而降低,至约600例TJA(OR=1.2),但更高年医院手术量未见进一步风险降低。
临床怎么用
- 原文结论指出模型预测性能差,不适于临床使用。
- 原文结论认为手术量阈值可能间接有助于指导政策及患者选择医院或外科医生。
局限性
- 原文未报告局限性。
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: Unplanned hospital readmissions after total joint arthroplasty (TJA) represent potentially serious adverse events and remain a critical measure of hospital quality. Predicting the risk of readmission after TJA may provide patients and clinicians with valuable information for preoperative decision-making. QUESTIONS/PURPOSES: (1) Can nonlinear machine-learning models integrating preoperatively available patient, surgeon, hospital, and county-level information predict 30-day unplanned hospital readmissions in a large cohort of nationwide Medicare beneficiaries undergoing TJA? (2) Which predictors are the most important in predicting 30-day unplanned hospital readmissions? (3) What specific information regarding population-level associations can we obtain from interpreting partial dependency plots (plots describing, given our modeling choice, the potentially nonlinear shape of associations between predictors and readmissions) of the most important predictors of 30-day readmission? METHODS: National Medicare claims data (chosen because this database represents a large proportion of patients undergoing TJA annually) were analyzed for patients undergoing inpatient TJA between October 2016 and September 2018. A total of 679,041 TJAs (239,391 THAs [61.3% women, 91.9% White, 52.6% between 70 and 79 years old] and 439,650 TKAs [63.3% women, 90% White, 55.2% between 70 and 79 years old]) were included. Model features included demographics, county-level social determinants of health, prior-year (365-day) hospital and surgeon TJA procedure volumes, and clinical classification software-refined diagnosis and procedure categories summarizing each patient's Medicare claims 365 days before TJA. Machine-learning models, namely generalized additive models with pairwise interactions (prediction models consisting of both univariate predictions and pairwise interaction terms that allow for nonlinear effects), were trained and evaluated for predictive performance using area under the receiver operating characteristic (AUROC; 1.0 = perfect discrimination, 0.5 = no better than random chance) and precision-recall curves (AUPRC; equivalent to the average positive predictive value, which does not give credit for guessing "no readmission" when this is true most of the time, interpretable relative to the base rate of readmissions) on two holdout samples. All admissions (except the last 2 months' worth) were collected and split randomly 80%/20%. The training cohort was formed with the random 80% sample, which was downsampled (so it included all readmissions and a random, equal number of nonreadmissions). The random 20% sample served as the first test cohort ("random holdout"). The last 2 months of admissions (originally held aside) served as the second test cohort ("2-month holdout"). Finally, feature importances (the degree to which each variable contributed to the predictions) and partial dependency plots were investigated to answer the second and third research questions. RESULTS: For the random holdout sample, model performance values in terms of AUROC and AUPRC were 0.65 and 0.087, respectively, for THA and 0.66 and 0.077, respectively, for TKA. For the 2-month holdout sample, these numbers were 0.66 and 0.087 and 0.65 and 0.075. Thus, our nonlinear models incorporating a wide variety of preoperative features from Medicare claims data could not well-predict the individual likelihood of readmissions (that is, the models performed poorly and are not appropriate for clinical use). The most predictive features (in terms of mean absolute scores) and their partial dependency graphs still confer information about population-level associations with increased risk of readmission, namely with older patient age, low prior 365-day surgeon and hospital TJA procedure volumes, being a man, patient history of cardiac diagnoses and lack of oncologic diagnoses, and higher county-level rates of hospitalizations for ambulatory-care sensitive conditions. Further inspection of partial dependency plots revealed nonlinear population-level associations specifically for surgeon and hospital procedure volumes. The readmission risk for THA and TKA decreased as surgeons performed more procedures in the prior 365 days, up to approximately 75 TJAs (odds ratio [OR] = 1.2 for TKA and 1.3 for THA), but no further risk reduction was observed for higher annual surgeon procedure volumes. For THA, the readmission risk decreased as hospitals performed more procedures, up to approximately 600 TJAs (OR = 1.2), but no further risk reduction was observed for higher annual hospital procedure volumes. CONCLUSION: A large dataset of Medicare claims and machine learning were inadequate to provide a clinically useful individual prediction model for 30-day unplanned readmissions after TKA or THA, suggesting that other factors that are not routinely collected in claims databases are needed for predicting readmissions. Nonlinear population-level associations between low surgeon and hospital procedure volumes and increased readmission risk were identified, including specific volume thresholds above which the readmission risk no longer decreases, which may still be indirectly clinically useful in guiding policy as well as patient decision-making when selecting a hospital or surgeon for treatment. LEVEL OF EVIDENCE: Level III, therapeutic study.
引用
- Kunze KN, So MM, Padgett DE, et al. Machine Learning on Medicare Claims Poorly Predicts the Individual Risk of 30-Day Unplanned Readmission After Total Joint Arthroplasty, Yet Uncovers Interesting Population-level Associations With Annual Procedure Volumes. Clin Orthop Relat Res. 2023. PMID: 37256278. doi:10.1097/CORR.0000000000002705
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。