Sebastian Johannes Fritsch
◽
Konstantin Sharafutdinov
◽
Moein Einollahzadeh Samadi
◽
Gernot Marx
◽
Andreas Schuppert
◽
...
BACKGROUND
During the course of the COVID-19 pandemic, a variety of machine learning models were developed to predict different aspects of the disease, such as long-term causes, organ dysfunction or ICU mortality. The number of training datasets used has increased significantly over time. However, these data now come from different waves of the pandemic, not always addressing the same therapeutic approaches over time as well as changing outcomes between two waves. The impact of these changes on model development has not yet been studied.
OBJECTIVE
The aim of the investigation was to examine the predictive performance of several models trained with data from one wave predicting the second wave´s data and the impact of a pooling of these data sets. Finally, a method for comparison of different datasets for heterogeneity is introduced.
METHODS
We used two datasets from wave one and two to develop several predictive models for mortality of the patients. Four classification algorithms were used: logistic regression (LR), support vector machine (SVM), random forest classifier (RF) and AdaBoost classifier (ADA). We also performed a mutual prediction on the data of that wave which was not used for training. Then, we compared the performance of models when a pooled dataset from two waves was used. The populations from the different waves were checked for heterogeneity using a convex hull analysis.
RESULTS
63 patients from wave one (03-06/2020) and 54 from wave two (08/2020-01/2021) were evaluated. For both waves separately, we found models reaching sufficient accuracies up to 0.79 AUROC (95%-CI 0.76-0.81) for SVM on the first wave and up 0.88 AUROC (95%-CI 0.86-0.89) for RF on the second wave. After the pooling of the data, the AUROC decreased relevantly. In the mutual prediction, models trained on second wave´s data showed, when applied on first wave´s data, a good prediction for non-survivors but an insufficient classification for survivors. The opposite situation (training: first wave, test: second wave) revealed the inverse behaviour with models correctly classifying survivors and incorrectly predicting non-survivors. The convex hull analysis for the first and second wave populations showed a more inhomogeneous distribution of underlying data when compared to randomly selected sets of patients of the same size.
CONCLUSIONS
Our work demonstrates that a larger dataset is not a universal solution to all machine learning problems in clinical settings. Rather, it shows that inhomogeneous data used to develop models can lead to serious problems. With the convex hull analysis, we offer a solution for this problem. The outcome of such an analysis can raise concerns if the pooling of different datasets would cause inhomogeneous patterns preventing a better predictive performance.
Model development and prediction of anti-icing longevity of asphalt pavement with salt-storage additive
