Prediction and recommendation by machine learning through repetitive internal validation for hepatic veno-occlusive disease/sinusoidal obstruction syndrome and early death after allogeneic hematopoietic cell transplantation

Constructing training set using distance between learnt graphical models of time series data on patient physiology, to predict disease scores

“…The diverse machine learning tools that [ 21 ] use to perform automated prediction in test cases, include logistic regression; Naïve Bayes; bagging techniques such as Random Forest; and boosting techniques such as Extreme gradient boosting (XGBoost) and Adaboost. Of these, XGBoost was noted to achieve optimal accuracy of prediction.…”

“…In lieu of an objective score for VOD onset and development, it may appear possible to undertake the learning of the pattern in the data collected from a retrospective set of bone-marrow transplant recipients, whose sufferance of VOD has been identified using a given model or interpretation of severity categorisation. We will need to review such attempts at using standard machine learning (ML) tools, [ 21 ], in the context of the question that Haematologist-Oncologists’ fundamentally desire an answer to, at the pre-transplant stage, namely: “how severely will VOD develop in a prospective patient?”. The efficacy of the reported ML tools in providing a reliable and explainable (and preferably continuous-valued) quantification of the virulence of VOD progression in the prospective patient will need to be reviewed, against any “black boxed” facility that lacks interpretability; lacks generalisability—to other implementation of VOD mitigation and HSCT protocol parameters—in spite of the large sample size, (as flagged up in the critical review by [ 22 ]).…”

“…2 Our work in the context of existing work [21] report on the results of the prediction of VOD status in a prospective recipient of HSCT, by (supervised) learning of the pattern in the data comprising observations on 20 selected features (as the inputs), and the severity category (as a target variable), using a variety of machine learning tools, for three different models or definitions that they invoke to assess VOD status in each of about 2500 HSCT recipients in a retrospective set. These models respectively ask: whether VOD did onset in a patient in this retrospective set or not; whether severe-to-very severe VOD happened to any such patient or not; and whether "early death" was noted in a patient or not.…”