One of the primary challenges we encountered was a significant class imbalance, with a higher number of patients withdrawing from treatment compared to those staying.
To address this issue, we implemented four different training approaches or pipelines on both the pre-pandemic and post-pandemic training datasets:
1.**Using the Original Dataset**: The models were trained on the original datasets.
2.**Class Weight Adjustment**: The models were trained on the original datasets but were penalized more heavily for misclassifying the minority class.
3.**Oversampling**: Additional samples were generated for the minority class (patients staying) to balance the dataset.
4.**Undersampling**: Samples from the majority class (patients withdrawing) were reduced to achieve balance.
These approaches resulted in multiple training datasets. However, to ensure a fair comparison of the models' performance across different pipelines, we utilized a common test dataset for evaluation, irrespective of the training approach followed.
