A Multi-Stage Machine Learning Model for Early Prediction of Neuro-logical Complications in Dengue Patients

Abstract

Dengue fever is still a major public health problem, especially in tropical and subtropical areas. Neurological complications are now being reported more often in dengue patients, but detecting them early is still not very easy. In this study, we tried to develop a multi-stage machine learning model to predict these complications early using routine clinical and lab parameters which are easily available. The approach is based on two stages. In stage 1, the model checks if there is any chance of neurological complication in the patient. Then in stage 2, it tries to predict specific conditions like neck rigidity, papilloedema and seizures. The XGBoost model performed best in stage 1, giving an AUC of 0.9286 and F1-score of 0.9167, which shows quite good performance, although not perfect. Some additional steps like handling class imbalance, probability calibration and selecting a proper threshold were also used, so that the model can work better in real clinical settings. The results are fairly good, especially for predicting neck rigidity and seizures, but still there can be some variation

depending on data.

One advantage is that the model is non-invasive, so it may help doctors in early decision making, mainly in places where resources are limited or advanced tests are not easily available. In future, the model should be tested on external datasets and maybe integrated into clinical decision systems for practical use.