An Evolutionary Approach to Interpretable Machine Learning for ICU Length of Stay Prediction

Neural networks have repeatedly demonstrated the ability to accurately model complex nonlinear relationships. However, the black-box nature of such models has led to concerns about bias and trustworthiness especially in settings like healthcare where it impacts multiple aspects of patient care including prognosis and resource-allocation. Existing model-agnostic methods to enhance neural network interpretability only offer an external perspective on the working of neural networks. To solve this issue, we designed an Evolutionary Multiple Gaussian Model that replaces the activation function of a perceptron with an aggregate of multiple Gaussian functions. The parameters of this model were evaluated using the gravitational search algorithm with the objective of minimizing a function of the mean absolute percentage error (MAPE) and mean squared error between the predicted and actual values. We performed multiple experiments on the Medical Information Mart for Intensive Care-IV dataset to predict the patient length of stay in Intensive Care Units. Results indicate that the proposed algorithm outperforms it's popular counterparts in terms of MAPE while offering better interpretability.