Use of What-if Scenarios to Help Explain Artificial Intelligence Models for Neonatal Health

TL;DR

A deep learning framework that not only predicts adverse labor outcomes from maternal, fetal, obstetrical, and intrapartum risk factors but also provides the model's reasoning behind the predictions made, and addresses the challenges of imbalance and small datasets.

Abstract

Early detection of intrapartum risk enables interventions to potentially prevent or mitigate adverse labor outcomes such as cerebral palsy. Currently, there is no accurate automated system to predict such events to assist with clinical decision-making. To fill this gap, we propose "Artificial Intelligence (AI) for Modeling and Explaining Neonatal Health" (AIMEN), a deep learning framework that not only predicts adverse labor outcomes from maternal, fetal, obstetrical, and intrapartum risk factors but also provides the model's reasoning behind the predictions made. The latter can provide insights into what modifications in the input variables of the model could have changed the predicted outcome. We address the challenges of imbalance and small datasets by synthesizing additional training data using Adaptive Synthetic Sampling (ADASYN) and Conditional Tabular Generative Adversarial Networks (CTGAN). AIMEN uses an ensemble of fully-connected neural networks as the backbone for its classification with the data augmentation supported by either ADASYN or CTGAN. AIMEN, supported by CTGAN, outperforms AIMEN supported by ADASYN in classification. AIMEN can predict a high risk for adverse labor outcomes with an average F1 score of 0.784. It also provides counterfactual explanations that can be achieved by changing 2 to 3 attributes on average. Resources available: https://github.com/ab9mamun/AIMEN.

Figures (8)

• Figure 1: AIMEN uses 34 risk factors of four categories. A machine learning model is trained and used to infer the risk of cerebral palsy. It also provides counterfactual explanations of the decision made The descriptions of the risk factors can be found in this paper mamun2023neonatal.
• Figure 2: The AIMEN system is made of three major components: a data generator, a risk predictor, and a risk explainer. These three components allow AIMEN to learn from small and challenging datasets and provide useful explanations of a prediction or diagnosis.
• Figure 3: AIMEN's backbone is an ensemble of eight fully connected neural networks. The default AIMEN has a specific backbone for the classifiers, that is MLP_v5. Eight neural networks of the same architecture (e.g. MLP_v5) are trained and validated on eight different folds of the cross-validation, and weighted voting among those eight models is performed through the ensemble network to classify on the test set.
• Figure 4: Performance metrics on the test set using different methods of data generation. The real training data features have only positive integers, however, generated data can have fractional and out-of-range (negative) values by default. Float means the classifiers were trained with fractional values and Int means generated data were converted to integers before training the classifiers. F1+ is the F1 score of the abnormal class, F1- is the F1 score of the normal class, and mF1 is the macro average F1 score. +, - in the legend represents both positive and negative values were present in the generated training data, whereas, + means all the values of the training data were positive.
• Figure 5: Training, validation, and test set metrics along with the test set confusion matrix with the AIMEN system with CTGAN data augmentation tool and MLP v5 backbone.