Ulivi Michele, Meroni Valentina, Orlandini Luca, Prandoni Lorenzo, Rossi Nicolò, Peretti Giuseppe M, Dui Linda Greta, Mangiavini Laura, Ferrante Simona
IRCCS Istituto Ortopedico Galeazzi, Via Riccardo Galeazzi, 4, 20161, Milan, Italy.
Residency Programme in Orthopedics and Traumatology, University of Milan, Via Festa del Perdono 7, 20122, Milan, Italy.
Comput Biol Med. 2020 Jun;121:103775. doi: 10.1016/j.compbiomed.2020.103775. Epub 2020 May 4.
Clinical registries are powerful tools for collecting uniform data longitudinally, thus making it possible to evaluate the outcome of patients affected by a specific pathology. In the context of total joint arthroplasty, registries serve also as post-market surveillance. Adoption of registries is a heavy burden for clinical settings in terms of resources and infrastructures. Excessive workload leads to incomplete data collection which undermines the effectiveness of a registry and consequently the workload needs to be optimised.
Starting from the use case of the Istituto Ortopedico Galeazzi, the time and personnel dedicated to the registry was estimated. Analysis of the data collected in the first years enabled us to propose a methodology for workload reduction. Different Machine Learning models were leveraged to predict patients with excellent satisfaction to reduce the number of assessments in their clinical post-operative follow-up. Moreover, feature selection was used to identify any unnecessary clinical scale to collect.
Given an acceptance rate of 3500 patients per year, 22 doctors and 6 non-medical employees were required to adopt a registry properly. Among the tested models, the Naïve Bayes gave the best performance (AUPRC = 0.81) in predicting patient satisfaction at six months. Moreover, we found that the 12-item Short Form was poorly informative in predicting satisfaction at six-months.
In this study machine learning was leveraged to provide a methodology to reduce workload in the use of pathology registries. Such workload reduction can have a considerable impact at a larger scale, and improve registry feasibility in high-volume hospitals.
临床注册是纵向收集统一数据的有力工具，因此能够评估受特定疾病影响患者的治疗结果。在全关节置换的背景下，注册还可作为上市后监测。对于临床机构而言，采用注册在资源和基础设施方面是一项沉重负担。工作量过大导致数据收集不完整，这会削弱注册的有效性，因此需要优化工作量。
从加莱阿齐骨科研究所的用例出发，估算了用于注册的时间和人员。对最初几年收集的数据进行分析，使我们能够提出一种减少工作量的方法。利用不同的机器学习模型预测满意度高的患者，以减少其术后临床随访中的评估次数。此外，还使用特征选择来确定任何不必要收集的临床量表。
若每年接受3500名患者，恰当地采用注册需要22名医生和6名非医疗人员。在测试的模型中，朴素贝叶斯在预测六个月时的患者满意度方面表现最佳（精确率均值下的面积 = 0.81）。此外，我们发现12项简短量表在预测六个月时的满意度方面信息量不足。
在本研究中，利用机器学习提供了一种减少疾病注册使用中工作量的方法。这种工作量的减少在更大范围内可能会产生相当大的影响，并提高大型医院注册的可行性。
