Explaining predictive factors in patient pathways using autoencoders

This paper introduces an end-to-end methodology to predict a pathway-related outcome and identifying predictive factors using autoencoders. A formal description of autoencoders for explainable binary predictions is presented, along with two objective functions that allows for filtering and inverting negative examples during training. A methodology to model and transform complex medical event logs is also proposed, which keeps the pathway information in terms of events and time, as well as the hierarchy information carried in medical codes. A case study is presented, in which the short-term mortality after the implementation of an Implantable Cardioverter-Defibrillator is predicted. Proposed methodologies have been tested and compared to other predictive methods, both explainable and not explainable. Results show the competitiveness of the method in terms of performances, particularly the use of a Variational Auto Encoder with an inverse objective function. Finally, the explainability of the method has been demonstrated, allowing for the identification of interesting predictive factors validated using relative risks.

Thank you for highlighting great opportunities for us to improve the manuscript, and for the very detailed suggestions. Please find bellow a point-by-point answer to all the recommendations.
"1) Section 3 -line 202 (below eqn 8): The meaning of the sentence " The explanation element…that the any input elements…" is not clear due to apparent mistake in sentence structure. Consider rewriting." The structure of the sentence has been addressed.
"2) Section 5.2.1 -line 322: The acronym AUC-ROC was not defined, but was only defined in a later Section 5.2.3 (line 360)." Definition has been added.
"3) Section 5.2.2: The authors state the selection of activation functions, layer dimension and loss function. Any particular reason why these hyperparameters were chosen? Was there any analysis/ hyperparameter tuning being made to optimise these hyperparameters, particularly because they are the same for all five DL models?" It is true that no hyperparameter tunning has been done to select the architecture of deep learning models. The activation function has been chosen as state-of-the-art choice for binary classification. Layer dimensions have been selected to have similar model size to compare with the encoder. Regarding activation function, hyperbolic tangent has been selected as a widespread choice for recurrent networks.
"5) Section 5.2.3 -line 365: "DT, RF and LR…. [43]. Deep learning… [44]". These sentences are general statements (not specific to the autoencoder section). It is suggested that a separate section (5.2.4) be added to include these statements as to provide clarity and avoid confusion to the readers. The last sentence of Section 5.2.3 "The experiments…. Windows 10 OS" could also be shifted to this newly added section." Section 5.2.4 added (Computational details).
"6) Section 5.3 and Fig 5: The authors present quantitative results, including values for the AUC-ROC, AUC-PR, and MCC. The reviewer suggests that the authors give a simple explanation of these results, particularly what these values represent in relation to the case study presented? Also, any particular reason of the discrepancy in results between AE(J_I) and AE(J_F)? This should also be included in the discussion." A description has been added in Section 5.3 in order to comment the difference obtained between AE_JI vs AE_JF. The methodology presented here is in fact a specific methodology based on the results obtained for that particular case study (i.e. the obtention of two categories of patterns -frequent and last time window -with for each pattern, a measure of the relative risk). The section has been reorganized in order to better highlight the specificity of the described methodology to that particular case study.
"10) Conclusion -line 456: The authors mentioned that better strategies can be used to improve prediction performance and explainability. Could the authors provide some possible examples/potential methods to achieve this? If not, this sentence remains highly speculative." This is indeed too speculative, investigations are still part of future work. The sentence has been rephrased.
"11) This study focuses exclusively on patient pathways to explain predictive factors, while omitting patient characteristics such as sex, age, and race. Would the inclusion of these characteristics in the study improve performance and/or uncover hidden patterns? How would the omission of these variables affect potential clinical applications? As the mortality or other patient outcomes could be affected by such factors , ie there is an inherent predisposition of the patient due to these factors." This was only slightly addressed in the conclusion. More details have been provided about how to extend the methodology to include patient characteristics.