Stroke is a significant cause of mortality and morbidity, necessitating early predictive strategies to minimize risks. Traditional methods for evaluating patients, such as Acute Physiology and Chronic Health Evaluation (APACHE II, IV) and Simplified Acute Physiology Score III (SAPS III), have limited accuracy and interpretability. This paper proposes a novel approach: an interpretable, attention-based transformer model for early stroke mortality prediction. This model seeks to address the limitations of previous predictive models, providing both interpretability (providing clear, understandable explanations of the model) and fidelity (giving a truthful explanation of the model\'s dynamics from input to output). Furthermore, the study explores and compares fidelity and interpretability scores using Shapley values and attention-based scores to improve model explainability. The research objectives include designing an interpretable attention-based transformer model, evaluating its performance compared to existing models, and providing feature importance derived from the model.

Explainability is key to enhancing the trustworthiness of artificial intelligence in medicine. However, there exists a significant gap between physicians\' expectations for model explainability and the actual behavior of these models. This gap arises from the absence of a consensus on a physician-centered evaluation framework, which is needed to quantitatively assess the practical benefits that effective explainability should offer practitioners. Here, we hypothesize that superior attention maps, as a mechanism of model explanation, should align with the information that physicians focus on, potentially reducing prediction uncertainty and increasing model reliability. We employed a multimodal transformer to predict lymph node metastasis of rectal cancer using clinical data and magnetic resonance imaging. We explored how well attention maps, visualized through a state-of-the-art technique, can achieve agreement with physician understanding. Subsequently, we compared two distinct approaches for estimating uncertainty: a standalone estimation using only the variance of prediction probability, and a human-in-the-loop estimation that considers both the variance of prediction probability and the quantified agreement. Our findings revealed no significant advantage of the human-in-the-loop approach over the standalone one. In conclusion, this case study did not confirm the anticipated benefit of the explanation in enhancing model reliability. Superficial explanations could do more harm than good by misleading physicians into relying on uncertain predictions, suggesting that the current state of attention mechanisms should not be overestimated in the context of model explainability.

BACKGROUND: Adverse events refer to incidents with potential or actual harm to patients in hospitals. These events are typically documented through patient safety event (PSE) reports, which consist of detailed narratives providing contextual information on the occurrences. Accurate classification of PSE reports is crucial for patient safety monitoring. However, this process faces challenges due to inconsistencies in classifications and the sheer volume of reports. Recent advancements in text representation, particularly contextual text representation derived from transformer-based language models, offer a promising solution for more precise PSE report classification. Integrating the machine learning (ML) classifier necessitates a balance between human expertise and artificial intelligence (AI). Central to this integration is the concept of explainability, which is crucial for building trust and ensuring effective human-AI collaboration.
OBJECTIVE: This study aims to investigate the efficacy of ML classifiers trained using contextual text representation in automatically classifying PSE reports. Furthermore, the study presents an interface that integrates the ML classifier with the explainability technique to facilitate human-AI collaboration for PSE report classification.
METHODS: This study used a data set of 861 PSE reports from a large academic hospital\'s maternity units in the Southeastern United States. Various ML classifiers were trained with both static and contextual text representations of PSE reports. The trained ML classifiers were evaluated with multiclass classification metrics and the confusion matrix. The local interpretable model-agnostic explanations (LIME) technique was used to provide the rationale for the ML classifier\'s predictions. An interface that integrates the ML classifier with the LIME technique was designed for incident reporting systems.
RESULTS: The top-performing classifier using contextual representation was able to obtain an accuracy of 75.4% (95/126) compared to an accuracy of 66.7% (84/126) by the top-performing classifier trained using static text representation. A PSE reporting interface has been designed to facilitate human-AI collaboration in PSE report classification. In this design, the ML classifier recommends the top 2 most probable event types, along with the explanations for the prediction, enabling PSE reporters and patient safety analysts to choose the most suitable one. The LIME technique showed that the classifier occasionally relies on arbitrary words for classification, emphasizing the necessity of human oversight.
CONCLUSIONS: This study demonstrates that training ML classifiers with contextual text representations can significantly enhance the accuracy of PSE report classification. The interface designed in this study lays the foundation for human-AI collaboration in the classification of PSE reports. The insights gained from this research enhance the decision-making process in PSE report classification, enabling hospitals to more efficiently identify potential risks and hazards and enabling patient safety analysts to take timely actions to prevent patient harm.

In recent years, the availability of publicly available data related to public health has significantly increased. These data have substantial potential to develop public health policy; however, this requires meaningful and insightful analysis. Our aim is to demonstrate how data analysis techniques can be used to address the issues of data reduction, prediction and explanation using online available public health data, in order to provide a sound basis for informing public health policy.
Observational suicide prevention data were analysed from an existing online United Kingdom national public health database. Multi-collinearity analysis and principal-component analysis were used to reduce correlated data, followed by regression analyses for prediction and explanation of suicide.
Multi-collinearity analysis was effective in reducing the indicator set of predictors by 30% and principal component analysis further reduced the set by 86%. Regression for prediction identified four significant indicator predictors of suicide behaviour (emergency hospital admissions for intentional self-harm, children leaving care, statutory homelessness and self-reported well-being/low happiness) and two main component predictors (relatedness dysfunction, and behavioural problems and mental illness). Regression for explanation identified significant moderation of a well-being predictor (low happiness) of suicide behaviour by a social factor (living alone), thereby supporting existing theory and providing insight beyond the results of regression for prediction. Two independent predictors capturing relatedness needs in social care service delivery were also identified.
We demonstrate the effectiveness of regression techniques in the analysis of online public health data. Regression analysis for prediction and explanation can both be appropriate for public health data analysis for a better understanding of public health outcomes. It is therefore essential to clarify the aim of the analysis (prediction accuracy or theory development) as a basis for choosing the most appropriate model. We apply these techniques to the analysis of suicide data; however, we argue that the analysis presented in this study should be applied to datasets across public health in order to improve the quality of health policy recommendations.