Comparison with Cox model’s performance. The Xxx proportional hazards regression model assumes proportionality of hazards across different prognostic groups over time. Any interaction between predictors and/or time needs to be manually specified by the user (e.g., fractional polynomials, splines). This may be difficult when a large set of prognostic factors is available. ML techniques such as ANNs, which are flexible and data-adaptive, relax this assumption and can naturally incorporate multi-way interactions between the input features. This characteristic together with the rise of computational power and the collection of large-volumes of data (with electronic healthcare records) has contributed to the popularity of XXXx. However, the Cox model remains the most common choice for survival data. Therefore, any new prediction model including SNNs should be compared to the traditional Cox model to be considered in clinical practice. Of 24 studies, 19 reported comparisons between Xxx models and SNNs. We assessed whether interaction terms were specified in the models to obtain optimal predictive performance in Xxx regression. Fifteen studies (78.9%) did not consider interaction terms between the predictors, information was unclear for 2 studies (10.5%), and 2 simulation studies considered interaction terms when applicable (10.5%). This result suggests suboptimal attention to the development of Xxx models, which in turn undermines inferences made regarding comparative SNN and Cox model performance. For datasets with a large number of prognostic factors (p > 10), a number of interaction terms can be selected based on external knowledge and clinical expertise (see [6]). Secondly, the author’s claim for the performance of SNN was investigated. Among the 19 studies comparing SNN and Cox model’s performance, 9 (47.4%) claimed better predictive performance of the SNN, while 5 reported a similar or better performance (26.3%) of the SNN compared to the Cox model. The performance was similar to Xxx’x model in 5 studies (26.3%). These result may be influenced by publication bias, as articles with favorable results are more likely to be published than articles with poor results. A fair comparison between SNN and Cox model approaches to modelling survival data should include model val- idation with proper evaluation measures, a comparison of calibration curves and the inclusion of non-linear terms and interactions for Xxx models, where applicable and possible. On the preface of his textbook on clinical predic- tio...
