Objectives

This is a protocol for a Cochrane Review (prognosis). The objectives are as follows:

To identify and evaluate AI‐based prognostic models, in development and external validation studies, which predict impending episodes of ventricular tachycardia from continuous surface electrocardiogram (ECG) recordings in individuals at elevated risk for ventricular arrhythmias due to any relevant underlying condition.

Purpose: Multimodal Machine Learning offers a holistic view of a patient’s status, integrating structured and unstructured data from electronic health records (EHR). We propose a framework to predict metastasis risk one month prior to diagnosis, using six months of clinical history from EHR data.

Methods: Data from four cancer cohorts collected at Karolinska University Hospital (Stockholm, Sweden) were analyzed: breast (n = 743), colon (n = 387), lung (n = 870), and prostate (n = 1890). The dataset included demographics, comorbidities, laboratory results, medications, and clinical text. We compared traditional and deep learning classifiers across single modalities and multimodal combinations, using various fusion strategies and a Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) 2a design, with an 80-20 development-validation split to ensure a rigorous, repeatable evaluation. Performance was evaluated using AUROC, AUPRC, F1 score, sensitivity, and specificity. We then employed a multimodal adaptation of SHAP to analyze the classifiers’ reasoning.

Results: Intermediate fusion achieved the highest F1 scores on breast (0.845), colon (0.786), and prostate cancer (0.845), demonstrating strong predictive performance. For lung cancer, the intermediate fusion achieved an F1 score of 0.819, while the text-only model achieved the highest, with an F1 score of 0.829. Deep learning classifiers consistently outperformed traditional models. Colon cancer, the smallest cohort, had the lowest performance, highlighting the importance of sufficient training data. SHAP analysis showed that the relative importance of modalities varied across cancer types.

Conclusion: Fusion strategies offer distinct strengths and weaknesses. Intermediate fusion consistently delivered the best results, but strategy choices should align with data characteristics and organizational needs.

Life-threatening conditions like sepsis are a leading cause of hospital mortality. The early identification of sepsis onset allows for timely intervention aiming to save patient lives. Although showing great promise for early sepsis onset prediction, Centralized Machine Learning applications are hindered by privacy concerns. Federated Learning has the potential to counteract the mentioned limitation as it trains a global model utilizing distributed data across several hospitals without sharing the data. This research explores the potential of Federated Learning to provide a more privacy-preserving and generalizable solution for predicting sepsis onset using a Deep Federated Learning setup. Patients from the MIMIC-III dataset are classified as either septic or non-septic using relevant patient features, and sepsis onset is identified at the first hour of a detected 5-hour SIRS interval for patients diagnosed with sepsis. We compare the predictive performance of different combinations of classifiers (LSTM and GRU), patient history window lengths, prediction window lengths, and Federated Learning clients, using the metrics AUROC, AUPRC, and F1-Score. Our results show that the Centralized Machine Learning and Federated Learning setups are on par in terms of predictive performance. In addition, on average, the best-performing Federated Learning model is GRU, with a five-hour patient history window and a three-hour prediction window. Overall, the study demonstrates that the proposed Federated Learning setup can predict sepsis onset comparably to state-of-the-art centralized deep learning algorithms for varying numbers of clients, enabling healthcare institutions to collaborate on mutually beneficial tasks without sharing isolated sensitive patient information.

Introduction: Asthma exacerbations remain a leading cause of morbidity and mortality despite effective treatments.

Aim: To evaluate if machine learning (ML) models can predict asthma exacerbations within one week using real-world home spirometry data.

Methods: A dataset of 467707 self-tests from 9599 Swedish asthma patients (2019–2024) collected via a home spirometry system was used. Each patient test comprised of demographics, lung function measurements, symptoms, triggers and treatment recommendations. Patients with at least five consecutive weekly measurements were included (n=3163). Asthma exacerbation was defined as a >20% decline in forced expiratory volume in one second (FEV₁) from patient’s personal best. Classical ML models were employed using repeated 5-week time-series windows.

Conclusions: The model identified over half of all exacerbations and 96% of non-events within one week. For future clinical use, encouraging more frequent spirometry testing in patients with undetected exacerbations may enhance model performance and support timely digital interventions to prevent exacerbations.

Federated Learning (FL) enables collaborative Machine Learning (ML) model training across institutions without sharing sensitive data, a critical advantage in healthcare. However, existing FL approaches for Intensive Care Unit (ICU) mortality prediction largely overlook explainability, which is essential for clinician trust. In this study, we present X4FL, a novel framework that integrates intrinsic and post-hoc interpretability into FL for ICU mortality prediction using the MIMIC-III dataset. X4FL employs an attention-based Long Short-Term Memory (LSTM) to capture temporal relevance at the feature level and combines it with SHapley Additive exPlanations (SHAP) to quantify feature importance. This dual explainability design yields clinically meaningful insights by highlighting both influential time segments and predictive features. Under a realistic non-Independent and Identically Distributed (non-IID) federated setup reflecting heterogeneous ICU populations, X4FL achieves predictive performance comparable to centralized models while consistently outperforming locally trained models, particularly in smaller or specialized ICUs. Attention heatmaps capture clinically relevant temporal patterns, and SHAP analyses reveal robust feature attributions across distributed settings. By addressing the trade-off between performance and interpretability, X4FL enhances transparency and trust in FL-based ICU mortality prediction, offering a step toward clinically viable federated solutions in critical care.

Applying deep learning models for healthcare-related forecasting applications has been widely adopted, such as leveraging glucose monitoring data of diabetes patients to predict hyperglycaemic or hypoglycaemic events. However, most deep learning models are considered black-boxes; hence, the model predictions are not interpretable and may not offer actionable insights into medical practitioners’ decisions. Previous work has shown that counterfactual explanations can be applied in forecasting tasks by suggesting counterfactual changes in time series inputs to achieve the desired forecasting outcome. This study proposes a generalized multivariate forecasting setup of counterfactual generation by introducing a novel approach, COMET, which imposes three domain-specific constraint mechanisms to provide counterfactual explanations for glucose forecasting. Moreover, we conduct the experimental evaluation using two diabetes patient datasets to demonstrate the effectiveness of our proposed approach in generating realistic counterfactual changes in comparison with a baseline approach. Our qualitative analysis evaluates examples to validate that the counterfactual samples are clinically relevant and can effectively lead the patients to achieve a normal range of predicted glucose levels by suggesting changes to the treatment variables.

Among recent developments in time series forecasting methods, deep forecasting models have gained popularity as they can utilize hidden feature patterns in time series to improve forecasting performance. Nevertheless, the majority of current deep forecasting models are opaque, hence making it challenging to interpret the results. While counterfactual explanations have been extensively employed as a post-hoc approach for explaining classification models, their application to forecasting models still remains underexplored. In this paper, we formulate the novel problem of counterfactual generation for time series forecasting, and propose an algorithm, called ForecastCF, that solves the problem by applying gradient-based perturbations to the original time series. The perturbations are further guided by imposing constraints to the forecasted values. We experimentally evaluate ForecastCF using four state-of-the-art deep model architectures and compare to two baselines. ForecastCF outperforms the baselines in terms of counterfactual validity and data manifold closeness, while generating meaningful and relevant counterfactuals for various forecasting tasks.

In machine learning applications, there is a need to obtain predictive models of high performance and, most importantly, to allow end-users and practitioners to understand and act on their predictions. One way to obtain such understanding is via counterfactuals, that provide sample-based explanations in the form of recommendations on which features need to be modified from a test example so that the classification outcome of a given classifier changes from an undesired outcome to a desired one. This paper focuses on the domain of time series classification, more specifically, on defining counterfactual explanations for univariate time series. We propose Glacier, a model-agnostic method for generating locally-constrained counterfactual explanations for time series classification using gradient search either on the original space or on a latent space that is learned through an auto-encoder. An additional flexibility of our method is the inclusion of constraints on the counterfactual generation process that favour applying changes to particular time series points or segments while discouraging changing others. The main purpose of these constraints is to ensure more reliable counterfactuals, while increasing the efficiency of the counterfactual generation process. Two particular types of constraints are considered, i.e., example-specific constraints and global constraints. We conduct extensive experiments on 40 datasets from the UCR archive, comparing different instantiations of Glacier against three competitors. Our findings suggest that Glacier outperforms the three competitors in terms of two common metrics for counterfactuals, i.e., proximity and compactness. Moreover, Glacier obtains comparable counterfactual validity compared to the best of the three competitors. Finally, when comparing the unconstrained variant of Glacier to the constraint-based variants, we conclude that the inclusion of example-specific and global constraints yields a good performance while demonstrating the trade-off between the different metrics.