Machine Learning and Artificial Intelligence (AI) more broadly have great immediate and future potential for transforming almost all aspects of medicine. However, in many applications, even outside medicine, a lack of transparency in AI applications has become increasingly problematic. This is particularly pronounced where users need to interpret the output of AI systems. Explainable AI (XAI) provides a rationale that allows users to understand why a system has produced a given output. The output can then be interpreted within a given context. One area that is in great need of XAI is that of Clinical Decision Support Systems (CDSSs). These systems support medical practitioners in their clinic decision-making and in the absence of explainability may lead to issues of under or over-reliance. Providing explanations for how recommendations are arrived at will allow practitioners to make more nuanced, and in some cases, life-saving decisions. The need for XAI in CDSS, and the medical field in general, is amplified by the need for ethical and fair decision-making and the fact that AI trained with historical data can be a reinforcement agent of historical actions and biases that should be uncovered. We performed a systematic literature review of work to-date in the application of XAI in CDSS. Tabular data processing XAI-enabled systems are the most common, while XAI-enabled CDSS for text analysis are the least common in literature. There is more interest in developers for the provision of local explanations, while there was almost a balance between post-hoc and ante-hoc explanations, as well as between model-specific and model-agnostic techniques. Studies reported benefits of the use of XAI such as the fact that it could enhance decision confidence for clinicians, or generate the hypothesis about causality, which ultimately leads to increased trustworthiness and acceptability of the system and potential for its incorporation in the clinical workflow. However, we found an overall distinct lack of application of XAI in the context of CDSS and, in particular, a lack of user studies exploring the needs of clinicians. We propose some guidelines for the implementation of XAI in CDSS and explore some opportunities, challenges, and future research needs.

Objective. Electroencephalography (EEG) is a key tool for non-invasive recording of brain activity and the diagnosis of epilepsy. EEG monitoring is also widely employed in rodent models to track epilepsy development and evaluate experimental therapies and interventions. Whereas automated seizure detection algorithms have been developed for clinical EEG, preclinical versions face challenges of inter-model differences and lack of EEG standardization, leaving researchers relying on time-consuming visual annotation of signals. Approach. In this study, a machine learning-based seizure detection approach, ‘Epi-AI’, which can semi-automate EEG analysis in multiple mouse models of epilepsy was developed. Twenty-six mice with a total EEG recording duration of 6451 h were used to develop and test the Epi-AI approach. EEG recordings were obtained from two mouse models of kainic acid-induced epilepsy (Models I and III), a genetic model of Dravet syndrome (Model II) and a pilocarpine mouse model of epilepsy (Model IV). The Epi-AI algorithm was compared against two threshold-based approaches for seizure detection, a local Teager-Kaiser energy operator (TKEO) approach and a global Teager-Kaiser energy operator-discrete wavelet transform (TKEO-DWT) combination approach. Main results. Epi-AI demonstrated a superior sensitivity, 91.4%–98.8%, and specificity, 93.1%–98.8%, in Models I–III, to both of the threshold-based approaches which performed well on individual mouse models but did not generalise well across models. The performance of the TKEO approach in Models I–III ranged from 66.9%–91.3% sensitivity and 60.8%–97.5% specificity to detect spontaneous seizures when compared with expert annotations. The sensitivity and specificity of the TKEO-DWT approach were marginally better than the TKEO approach in Models I–III at 73.2%–80.1% and 75.8%–98.1%, respectively. When tested on EEG from Model IV which was not used in developing the Epi-AI approach, Epi-AI was able to identify seizures with 76.3% sensitivity and 98.1% specificity. Significance. Epi-AI has the potential to provide fast, objective and reproducible semi-automated analysis of multiple types of seizure in long-duration EEG recordings in rodents.

In this study, a deep reinforcement learning approach is proposed to handle tactical driving in complex highway traffic environments for unmanned ground vehicles. Tactical driving is a challenging topic for unmanned ground vehicles because of its interplay with routing decisions as well as real-time traffic dynamics. The core of our deep reinforcement learning approach is a deep Q-network that takes dynamic traffic information as input and outputs typical tactical driving decisions as action. The reward is designed with the consideration of successful highway exit, average traveling speed, and driving safety and comfort. In order to endow an unmanned ground vehicle with situational traffic information that is critical for tactical driving, the vehicle’s sensor information such as vehicle position and velocity are further augmented through the assessment of the ego-vehicle’s collision risk, potential field, and kinematics and used as input for the deep Q-network model. A convolutional neural network is built and fine-tuned to extract traffic features which facilitate the decision-making process of Q-learning. For model training and testing, a highway simulation platform is constructed with realistic parameter settings obtained from a real-world highway traffic dataset. The performance of the deep Q-network model is validated with extensive simulation experiments under different parameter settings such as traffic density and risk level. The results exhibit the important potentials of our deep Q-network model in learning challenging tactical driving decisions given multiple objectives and complex traffic environment.

Objective: Sleep spindle features show developmental changes during infancy and have the potential to provide an early biomarker for abnormal brain maturation. Manual identification of sleep spindles in the electroencephalogram (EEG) is time-consuming and typically requires highly-trained experts. Automated detection of sleep spindles would greatly facilitate this analysis. Research on the automatic detection of sleep spindles in infant EEG has been limited to-date. Methods: We present a random forest-based sleep spindle detection method (Spindle-AI) to estimate the number and duration of sleep spindles in EEG collected from 141 ex-term born infants, recorded at 4 months of age. The signal on channel F4-C4 was split into a training set (81 ex-term) and a validation set (30 ex-term). An additional 30 ex-term infant EEGs (channel F4-C4 and channel F3-C3) were used as an independent test set. Fourteen features were selected for input into a random forest algorithm to estimate the number and duration of spindles and the results were compared against sleep spindles annotated by an experienced clinical physiologist. Results: The prediction of the number of sleep spindles in the independent test set demonstrated 93.3% to 93.9% sensitivity, 90.7% to 91.5% specificity, and 89.2% to 90.1% precision. The duration estimation of sleep spindle events in the independent test set showed a percent error of 5.7% to 7.4%. Conclusion and Significance: Spindle-AI has been implemented as a web server that has the potential to assist clinicians in the fast and accurate monitoring of sleep spindles in infant EEGs.

Amyotrophic Lateral Sclerosis (ALS), also known as Motor Neuron Disease (MND), is a rare and fatal neurodegenerative disease. As ALS is currently incurable, the aim of the treatment is mainly to alleviate symptoms and improve quality of life (QoL). We designed a prototype Clinical Decision Support System (CDSS) to alert clinicians when a person with ALS is experiencing low QoL in order to inform and personalise the support they receive. Explainability is important for the success of a CDSS and its acceptance by healthcare professionals. The aim of this work isto announce our prototype (C-ALS), supported by a first short evaluation of its explainability. Given the lack of similar studies and systems, this work is a valid proof-of-concept that will lead to future work. We developed a CDSS that was evaluated by members of the team of healthcare professionals that provide care to people with ALS in the ALS/MND Multidisciplinary Clinic in Dublin, Ireland. We conducted a user study where participants were asked to review the CDSS and complete a short survey with a focus on explainability. Healthcare professionals demonstrated some uncertainty in understanding the system’s output. Based on their feedback, we altered the explanation provided in the updated version of our CDSS. C-ALS provides local explanations of its predictions in a post-hoc manner, using SHAP (SHapley Additive exPlanations). The CDSS predicts the risk of low QoL in the form of a probability, a bar plot shows the feature importance for the specific prediction, along with some verbal guidelines on how to interpret the results. Additionally, we provide the option of a global explanation of the system’s function in the form of a bar plot showing the average importance of each feature. C-ALS is available online for academic use.

Breast cancer is one of the most prevalent cancers among women, with early detection playing a critical role in improving survival rates. This study introduces a novel transformer-based explainable model for breast cancer lesion segmentation (TEBLS), aimed at enhancing the accuracy and interpretability of breast cancer lesion segmentation in medical imaging. TEBLS integrates a multi-scale information fusion approach with a hierarchical vision transformer, capturing both local and global features by leveraging the self-attention mechanism. This model addresses the limitations of existing segmentation methods, such as the inability to effectively capture long-range dependencies and fine-grained semantic information. Additionally, TEBLS incorporates visualization techniques to provide insights into the segmentation process, enhancing the model’s interpretability for clinical use. Experiments demonstrate that TEBLS outperforms traditional and existing deep learning-based methods in segmenting complex breast cancer lesions with variations in size, shape, and texture, achieving a mean DSC of 81.86% and a mean AUC of 97.72% on the CBIS-DDSM test set. Our model not only improves segmentation accuracy but also offers a more explainable framework, which has the potential to be used in clinical settings.