Explainable Stacking-Based Model for Predicting Hospital Readmission for Diabetic Patients
Round 1
Reviewer 1 Report
Interesting work, I commend the authors on this excellent contribution. The methodology is largely clear, and the manuscript is well-written as well. It is also good to see the authors aware of possible limitations. I only have minor issues to consider in the next version, please.
(1)
Please elaborate further on the intuition behind the choice of stacking-based Machine Learning (ML). The reader would wonder why the authors preferred that approach over other state-of-the-art approaches, such as Deep Learning for example.
(2)
There is a dire need to refer to more up-to-date references in the related work. I recommend including more recent studies that applied the state-of-the-art ML for predicting patient admission/readmission. For example:
https://doi.org/10.3390/jpm10030082
https://doi.org/10.1109/BigData50022.2020.9378073
Author Response
Comment/Suggestion: 1
Please elaborate further on the intuition behind the choice of stacking-based Machine Learning (ML). The reader would wonder why the authors preferred that approach over other state-of-the-art approaches, such as the Deep Learning example.
Our response: Thank you for this comment. We added it to the revised manuscript. Please see lines 373 -383.
Comment/Suggestion: 2
There is a dire need to refer to more up-to-date references in the related work. I recommend including more recent studies that applied state-of-the-art ML for predicting patient admission/readmission.
Our response:
We have considered these four references in the revised manuscript. Please see lines 73 – 77 and the reference list for further details.
Reviewer 2 Report
I am really grateful for reviewing this manuscript. In my opinion, this manuscript has an issue of model performance. The area under the receiver-operating-characteristic curve of the stacked model in this study, 0.673, is almost the same with that of logistic regression in a previous study (0.670) in Table 4. Explainable artificial intelligence can be justified only when it has better performance than its previous counterparts.
Author Response
Comment/Suggestion: 1
I am really grateful for reviewing this manuscript. In my opinion, this manuscript has an issue of model performance. The area under the receiver-operating-characteristic curve of the stacked model in this study, 0.673, is almost the same as that of logistic regression in a previous study (0.670) in Table 4. Explainable artificial intelligence can be justified only when it performs better than its previous counterparts.
Our response:
Thanks for your feedback. The area under the receiver-operating-characteristic curve (AUC) of the logistic model in Table 4 comes from the training phase, but our AUC is from the testing phase. Training AUC is typically higher than testing because the model fits the training data. Please consider the comment column in Table 4. In addition, we have added further discussion in lines 373 – 383. We explained why we chose the stacking-based approach instead of other approaches (e.g., logistic regression and deep learning models).
Reviewer 3 Report
· The novelty of this paper is not clear. The difference between present work and previous Works should be highlighted.
· The author needs to change the abstract and focus more on problem domain. Before the paper contributions, the author could precisely include the need of developing the proposed method.
· How did the authors apply the Augmentation technique?
· The author could better explain how “Related works” is actually related to the current study. It is not clear to the reader how the manuscript is similar to or differs from these related works.
· Is there any limitation of the proposed work? If so, the author should include it at the end of the conclusion part. This may help future researchers to overcome the limitations.
Author Response
Comment/Suggestion: 1
The novelty of this paper is not clear. The difference between present work and previous Works should be highlighted.
Our response:
Thank you for this comment. We have highlighted it in lines 79 – 82 and 92 - 95. The existing works focus on the prediction of the risk of hospital readmission and classify high-risk factors. These studies often ignore the interpretability or lack of explainability of the models, leading to the ‘black box’ phenomenon. The novelty of this study is to use the stacking approach (a combination of state-of-the-art machine learning models) to predict 30-day readmission following the hospitalisation of diabetic patients using the XAI technique.
Comment/Suggestion: 2
The author needs to change the abstract and focus more on the problem domain. Before the paper’s contributions, the author could precisely include the need to develop the proposed method.
Our response:
We modified it in the revised manuscript. Please see lines 10-11 and 96-99.
Comment/Suggestion: 3
How did the authors apply the Augmentation technique?
Our response:
We applied our proposed framework to the US Health Facts Database, which comprised 101,766 de-identified diabetes patients between 1999 and 2008. Compared to the different baseline models, performance analysis shows our model can better predict readmission than other existing models. Also, our proposed model is explainable, we used two patients as a case study to explain the most important features and the prediction probabilities for readmission. Please see sections 2 and 3 for more details.
Comment/Suggestion: 4
The author could better explain how “Related works” is actually related to the current study. It is not clear to the reader how the manuscript is similar to or differs from these related works.
Our response:
We have explained it in lines 63 – 101. Our research uses a machine learning method to predict the risk of hospital readmission for diabetic patients, which is similar to the existing works. However, we proposed a framework with an explainable stacking-based model, which differs from these related works.
Comment/Suggestion: 5
Is there any limitation of the proposed work? If so, the author should include it at the end of the conclusion part. This may help future researchers to overcome the limitations.
Our response:
Yes, there are limitations. We have discussed this in lines 406 – 412.
Round 2
Reviewer 2 Report
I am really grateful for reviewing this manuscript. In my opinion, this manuscript can be published in current form.
Author Response
The reviewer is happy with our responses provided in the previous round
