Development of an Artificial Intelligence-Based Model for Patient’s Vital Signs Deterioration Prediction

This research focuses on establishing reliable machine learning models for clinical decision support by highlighting the crucial roles of data preprocessing and quality. Using the MIMIC-IV database, we developed and validated algorithms based on vital physiological indicators, including blood pressure, temperature, heart rate, respiratory rate, and blood oxygen saturation (SpO2).. The study reveals that the quality of the data provided to machine learning models significantly impacts their performance and reliability in clinical environments. To preserve data accuracy and quality, we have enforced rigorous data preprocessing and quality control guidelines, involving univariate and multivariate analyses. The refined data was utilized to educate an artificial neural network (ANN), which formulated an Early Warning Score (EWS) system. Remarkable performance was displayed by the model, with perfect classification scores (precision, recall, F1 score, and accuracy all equaling 1.0) for individual vital sign predictions. Additionally, the model's MSE and MAE were close to zero, indicating negligible error in the regression metrics. The AUC curve's area was consistently high across all parameters (ranging from 0.992 to 1.000), while the validation accuracy ranged from 94.6% to 100%. Such results are achievable when using high-quality data. Conversely, they also illustrate the negative effects of compromised data quality on performance. In conclusion, the successful development and trustworthy deployment of machine learning systems in healthcare settings rely on robust data preprocessing and quality control, as this research illustrates.