Review of 'Using machine learning for healthcare treatment planning'

This paper presents " a methodology for using machine learning for planning treatments ... As a case study, we apply the proposed methodology to Breast Cancer ... our paper focuses on applying Machine Learning to suggest treatment plans for patients with different disease severity." However, the model does not include the standard of care for breast cancer. Herceptin is not one of the treatments considered, nor are genetic tests considered to determine the cancer endotype.

If we set those issues aside, perhaps it can still be considered as a case study illustrating a more widely applicable method. 

They offer "an interactive model where a patient can enter his details and use the model to predict the probability of his staying alive based on a combination of Radiation Sequence, Radiation Recode, and Chemotherapy Recode." They don't seem to consider the horrific nature of this user experience. The peculiar use of the pronoun "he" in the sentence underlines this lack of empathy. The paper reports no ethical review, so presumably the authors met no patients, and conducted no User Experience trials.

This interactive model is based on a KNN model with just 25 neighbours, creating a large and unethical risk that the the results are misleading due to sampling error. They say "we believe that this would be a number that is sufficient to compute statistics on alternative treatments" but basic knowledge of sampling statistics would show that this is wrong.

This model is designed for "explaining and defending a particular treatment choice to the patient" - but that choice comes from a logistic regression. They offer no evidence that the explanation is consistent with the recommendation of the main model.

For both models, the predictand is 5 year survival, and then the model is used to choose treatments. However, more aggressive treatment is correlated with lower survival because it is applied when the disease is more serious. A model used to recommend treatment must incorporate all data about the patient's condition which is available to the clinician - without that the method will be biassed towards optimistic not realistic treatments.