Can machine-learning algorithms replace conventional statistics?

Abstract

Wong et al. 1 evaluate 19 clinical variables (training data) and three supervised machine-learning algorithms to predict early biochemical recurrence after robot-assisted prostatectomy. They further compare the areas under the curve (AUCs) resulting from these algorithms with the AUC of a conventional Cox regression model and conclude that the machine-learning algorithms can produce accurate disease prognosis, perhaps better than a traditional Cox regression model. As the authors state, predictive models have the potential to better individualize care to patients at highest risk of prostate cancer recurrence and progression. The authors should be commended for their adoption of machine-learning algorithms to better interpret the vast volumes of clinical data and assess prognosis after robot-assisted prostatectomy. This should represent another step forward for the management of prostate cancer, where tailored treatment is now largely based on the clinical risk stratification of the disease 2. Incidentally, we are also in an era where we are seeing aspects of artificial intelligence (machine learning being a subset of it) vastly transform how we view and process data in everyday life. This has been true in medicine as well, particularly for prostate cancer 3. While our own research group has also evaluated machine-learning algorithms to process surgeon performance metrics and predict clinical outcomes after robot-assisted prostatectomy 4, I want to express a word of caution. Utilization of machine learning does not in itself imply automatic superiority over conventional statistics 5 despite literature that has demonstrated so 3. The success of predictive models in machine learning still relies on the quality of data introduced and careful execution of the analysis. In our experience, it works best when highly experienced clinicians and data scientists are working hand in hand. Furthermore, I would argue that the results of this present study do not necessarily show that machine learning is superior to conventional statistics, but rather it highlights an inherent advantage of machine learning. While traditional analyses require the a priori selection of a model based on the available data, machine learning has more flexibility for model fitting 6. Additionally, inclusion of variables in traditional analyses is constrained by the sample size. In contrast, by design, machine learning models thrive on their ability to consider many variables concurrently, and as such, have the potential to detect underlying patterns that may otherwise be undetectable when data are examined effectively in individual silos. We look forward to the external validation of the methodology described in the present article. Big and diverse data are critical requirements of machine learning. A multi-institutional, multi-surgeon cohort is necessary to confirm the findings in this report. A further step from there is the adoption of such prediction models into clinical use. The ultimate question is how improved prognostic data may influence surgeon and patient decisions. Dr Hung reports personal fees from Ethicon, Inc, outside the submitted work.