Abstract
Automated healthcare decision support has seen a huge rise with the improved data collection models in hospitals and also improved machine learning based techniques that exhibit high possibilities for automation. Automating healthcare systems as an aide for clinical practitioners can ensure fast and more accurate results for the patients and can also aid in hospital administration. Mortality prediction has been one of the major and critical factors that determines the type of treatment and the level of resources that has to be allocated for a patient. This work presents a pipelined ensemble architecture that can be used for effective prediction of mortality levels of a patient. The pipeline model has been designed in multiple levels to ensure improvement of quality of the medical data and effective prediction. The pipelined architecture model has been compared with existing state-of-the-art model, and the results indicate high performance with 92% accuracy levels, ensuring the model is suitable for use in real time mortality prediction.
Introduction
Predicting patient mortality in Intensive Care Units (ICUs) is a critical task for effective triaging and resource allocation. However, clinical data is notoriously messy, sparse, and complex. Modern Electronic Health Records (EHRs) are massive and heterogeneous, containing data from various hospital departments that is often imbalanced—most patients are fortunately discharged alive. Manual interpretation of this data by clinicians in high-stakes environments is increasingly difficult, creating a strong need for automated, machine learning-driven clinical decision support systems.
This paper introduces the Pipelined Ensemble Architecture for Mortality Prediction (PEAMP). By combining specialized feature selection with an ensemble of neural networks and random forests, the PEAMP model addresses the inherent challenges of ICU data to provide reliable, real-time mortality predictions.
Methodology
The PEAMP architecture is a multi-stage pipeline designed to transform raw EHR data into actionable insights through several key phases:
- Data Preprocessing: The system merges disparate hospital department tables from the MIMIC III dataset, handles missing values, and converts categorical variables into numerical formats using one-hot encoding.
- Model-Based Feature Selection: To combat the "curse of dimensionality" caused by one-hot encoding, the authors utilize Logistic Regression as a meta-transformer. Features are weighted, and only those exceeding a mean weight threshold are retained for training.
- Multi-Model Integrated Training: The predictive engine employs a dual-model approach:
- Artificial Neural Network (ANN): A 100-neuron hidden layer using ReLU activation processes the full dataset to perform intrinsic feature selection.
- Random Forest: This model utilizes data subsets and tree pruning to specifically manage class imbalance and noise.
- Prediction Integration: A mean-based combiner aggregates the outputs from both the ANN and the Random Forest to produce the final mortality prediction.
Results and Discussion
The PEAMP model was evaluated using the Medical Information Mart for Intensive Care III (MIMIC III) dataset, which contains de-identified data from over 40,000 critical care patients. The model achieved an overall accuracy of approximately 92-93%.
When compared to the previous state-of-the-art model by Ding et al., PEAMP demonstrated significant improvements across nearly all primary metrics. Accuracy rose from 66% to 92%, sensitivity increased from 66.6% to 88%, and specificity saw a dramatic jump from 66.1% to 99%. Notably, the model achieved a Positive Predictive Value (PPV) of 99%, which is crucial in clinical settings to prevent "alarm fatigue" caused by false positives.
As shown in the detailed performance table, the model is particularly effective at identifying mortality (Class 1) with a precision of 1.0 and an F1-score of 0.94. While the sensitivity of 88% is considered moderate, it represents a substantial leap forward for automated ICU monitoring.
Conclusion
The PEAMP model demonstrates that a pipelined ensemble approach can effectively navigate the complexities of EHR data to provide high-accuracy mortality predictions. While the model currently shows a slight decrease in Negative Predictive Value (NPV) compared to some baselines, its near-perfect PPV makes it a strong candidate for real-time clinical decision support. Future work will focus on further improving sensitivity and optimizing the pipeline's parallel architecture for faster deployment in hospital environments.
Summary updated