Work

Interpretability of Deep Learning Models for Time-Series Clinical Data

Year: 2023
Type: dissertation
Abstract: Clinical decision-making concerning patients with traumatic brain injuries (TBI) requires accurate projections of both short- and long-term outcomes. However, models based on these data frequently lack transparency, making it difficult for physicians to trust and comprehend the decisions that are made using these models. Explainable Artificial Intelligence (XAI) has gained increasing interest in recent years as a viable method for enhancing the interpretability of these models. One of the reasons for this rise in interest is that XAI is a relatively novel subject.In this dissertation, we provide a comprehensive examination of the application of XAI algorithms to clinical time-series data in order to determine the prognosis of traumatic brain injury. (TBI). In the first study, a predictive model for traumatic brain injury (TBI) is constructed utilizing an attention-based recurrent neural network (RNN) that combines time-series data with other clinical characteristics. The model achieved a superior level of discriminative ability and outperformed other models. The inability of deep learning models to be interpreted by humans, however, poses a significant barrier to the widespread implementation of these techniques in medical settings. In the second study, XAI strategies for TBI prediction models are evaluated critically. This investigation compares the understandability, fidelity, and stability of six alternative methods. We developed two prediction models for the short-term and long-term outcomes of traumatic brain injury, using structured tabular data and time-series physiologic data, respectively. Our study demonstrates that there are numerous XAI techniques, each with its own set of advantages and disadvantages that researchers must consider when evaluating model decisions. In the third study, a novel framework called WindowSHAP is proposed to explain time-series classifiers using Shapley values, which are a well-known method for determining the significance of features. The framework's objective is to improve the overall quality of explanations while reducing the computational complexity of generating Shapley values for lengthy time-series data. Based on the results, it is evident that our framework is superior to other approaches in terms of both computing efficiency and interpretability. When taken as a whole, our findings provide light on the potential of XAI approaches to improve the interpretability and transparency of clinical time-series data models. The dissertation makes a substantial contribution to a deeper understanding of how XAI may be applied to clinical time-series data and provides insights into the advantages and disadvantages of various techniques. Our research has significant implications for the development of trustworthy prediction models for clinical decision-making in the context of traumatic brain injury (TBI) and other complicated medical disorders.
Authors: Nayebi Nodoushan, Amin, Nayebi Nodoushan, Amin
Language: English
Open Access status: green
OpenAlex ID: W7048006469
Cites: 60
Cited by:
Related to: 10