Editorial Comment on: Utilizing Machine Learning and Automated Performance Metrics to Evaluate Robot-Assisted Radical Prostatectomy Performance and Predict Outcomes by Hung et al.

The authors should be congratulated for their influential study evaluating the surgical performance and estimating clinical outcomes of robotic radical prostatectomy (RRP) operations by using “machine learning” (ML) algorithms.

ML has become a source of hugely exciting research and innovation in recent years. Its strength is in the ability to learn with data. Efficient learning algorithms have been developed in the past decades to handle “big data” and also to discover that hidden information. Nowadays it is called as “Machine learning” that can be categorized as supervised, unsupervised, and semisupervised algorithms. The first algorithm uses labeled data to teach itself, whereas the unsupervised algorithms try to describe a function form. Some of the techniques can use both the supervised and unsupervised algorithms.

In this study using supervised classification, techniques namely Random Forest algorithm, support vector machines, and logistic regression, L2 regularization is investigated to exhibit the most accurate algorithm by comparing the outputs with the actual values as the ground truth. The aim was acquiring the assessment of surgical performance that affects clinical outcomes.

These algorithms were trained using the automated performance metrics coming from the robot and with the hospital length of stay (LOS). Among the selected algorithms, “Random Forest–50” had the best performance for the selected automated performance metrics. Apparently, the hearth of the study is kinematic data included characteristics of movement such as instrument travel time, path length, and velocity. System events such as frequency of clutching, camera movements, third arm, and energy usage were also included. These data were used for predicting surgical performance and estimating clinical outcomes.

From the clinical point of view, a few perspectives should be emphasized. First of all, this study reveals a proof of concept for estimating the outcomes of RRP operations with the aid of digital operative data. With this model, surgery time, LOS, and urethral catheterization were predicted accurately. Even though some parameters such as LOS might be effected by several issues other than surgical experience such as comorbidities of the patients, this technology seems to be promising. Second, this model could be used for optimization of the patient care in the future. For example, ML algorithms may suggest the individualized surgical drainage or Foley catheter removal times, and this kind of management might be more cost-effective than a constant suggested removal times. Third, information obtained from ML algorithms might be incorporated with the anatomical knowledge derived from the imaging modalities such as multiparametric MRI for improving the functional and oncologic outcomes. And lastly, surgical training may be performed in a more scientific way instead of classical “see one, do one, teach one” principle. Certain cutoffs may be created using kinematic data for adequacy of the trainee before tackling the routine cases.

As a limitation, the data presented in this study were gathered from nine surgeons with different levels of expertise in RRP operations. Thus, data from ML algorithms might not represent the actual outcomes of individual surgeons. In contrast, the precise methodology of this study should have been used for estimating the functional and oncologic outcomes, which would reflect the real value of this article.