Deep learning has demonstrated expert-level performance in melanoma classification, positioning it as a powerful tool in clinical dermatology. However, model opacity and the lack of interpretability remain critical barriers to clinical adoption, as clinicians often struggle to trust the decision-making processes of black-box models. To address this gap, we present a Cross-modal Explainable Framework for Melanoma (CEFM) that leverages contrastive learning as the core mechanism for achieving interpretability. Specifically, CEFM maps clinical criteria for melanoma diagnosis—namely Asymmetry, Border, and Color (ABC)—into the Vision Transformer embedding space using dual projection heads, thereby aligning clinical semantics with visual features. The aligned representations are subsequently translated into structured textual explanations via natural language generation, creating a transparent link between raw image data and clinical interpretation. Experiments on public datasets demonstrate 92.79% accuracy and an AUC of 0.961, along with significant improvements across multiple interpretability metrics. Qualitative analyses further show that the spatial arrangement of the learned embeddings aligns with clinicians’ application of the ABC rule, effectively bridging the gap between high-performance classification and clinical trust.

Vital signs, such as heart rate and blood pressure, are critical indicators of patient health and are widely used in clinical monitoring and decision-making. While deep learning models have shown promise in forecasting these signals, their deployment in healthcare remains limited in part because clinicians must be able to trust and interpret model outputs. Without reliable uncertainty quantification - particularly calibrated prediction intervals (PIs) - it is unclear whether a forecasted abnormality constitutes a meaningful warning or merely reflects model noise, hindering clinical decision-making. To address this, we present two methods for deriving PIs from the Reconstruction Uncertainty Estimate (RUE), an uncertainty measure well-suited to vital-sign forecasting due to its sensitivity to data shifts and support for label-free calibration. Our parametric approach assumes that prediction errors and uncertainty estimates follow a Gaussian copula distribution, enabling closed-form PI computation. Our non-parametric approach, based on k-nearest neighbours (KNN), empirically estimates the conditional error distribution using similar validation instances. We evaluate these methods on two large public datasets with minute- and hour-level sampling, representing high- and low-frequency health signals. Experiments demonstrate that the Gaussian copula method consistently outperforms conformal prediction baselines on low-frequency data, while the KNN approach performs best on high-frequency data. These results underscore the clinical promise of RUE-derived PIs for delivering interpretable, uncertainty-aware vital sign forecasts.

Explaining machine learning (ML) predictions has become crucial as ML models are increasingly deployed in high-stakes domains such as healthcare. While SHapley Additive exPlanations (SHAP) is widely used for model interpretability, it fails to differentiate between causality and correlation, often misattributing feature importance when features are highly correlated. We propose Causal SHAP, a novel framework that integrates causal relationships into feature attribution while preserving many desirable properties of SHAP. By combining the Peter-Clark (PC) algorithm for causal discovery and the Intervention Calculus when the DAG is Absent (IDA) algorithm for causal strength quantification, our approach addresses the weakness of SHAP. Specifically, Causal SHAP reduces attribution scores for features that are merely correlated with the target, as validated through experiments on both synthetic and real-world datasets. This study contributes to the field of Explainable AI (XAI) by providing a practical framework for causal-aware model explanations. Our approach is particularly valuable in domains such as healthcare, where understanding true causal relationships is critical for informed decision-making.

In recent years, deep learning has gained traction in medical research, particularly in addressing challenges related to predicting outcomes from irregularly sampled time-series data. This irregularity, stemming from inconsistent follow-up appointments and vital sign recordings, poses significant hurdles for traditional machine learning techniques reliant on regularly sampled data. To overcome this, researchers have developed two main categories of methods: interpolation-based and non-interpolation-based models. Our study proposes STraTS-mTAND, a novel approach that integrates techniques from both categories to predict outcomes from irregularly sampled time-series data. By leveraging the strengths of STraTS, a non-interpolation model, and mTAND, an interpolation model, our approach aims to mitigate their respective limitations. Evaluation on datasets like the PhysioNet Challenge 2012 and MIMIC-III demonstrates superior performance compared to existing methods, showing improvements in both PR-AUC and ROC-AUC metrics. Notably, our approach exhibits robustness with less training data and performs effectively with sparser and more irregular time series—characteristics inherent in recorded medical data. Additionally, our analysis offers insights into optimizing predictive modelling in healthcare, potentially reducing costs associated with data extraction and management, shortening operational cycles for model development and deployment, and enhancing digital agility and sustainability within the healthcare sector.

The "similarity of dissimilarities" is an emerging paradigm in biomedical science with significant implications for protein function prediction, machine learning (ML), and personalized medicine. In protein function prediction, recognizing dissimilarities alongside similarities provides a more detailed understanding of evolutionary processes, allowing for a deeper exploration of regions that influence biological functionality. For ML models, incorporating dissimilarity measures helps avoid misleading results caused by highly correlated or similar data, addressing confounding issues like the Doppelgänger Effect. This leads to more accurate insights and a stronger understanding of complex biological systems. In the realm of personalized AI and precision medicine, the importance of dissimilarities is paramount. Personalized AI builds local models for each sample by identifying a network of neighboring samples. However, if the neighboring samples are too similar, it becomes difficult to identify factors critical to disease onset for the individual, limiting the effectiveness of personalized interventions or treatments. This paper discusses the "similarity of dissimilarities" concept, using protein function prediction, ML, and personalized AI as key examples. Integrating this approach into an analysis allows for the design of better, more meaningful experiments and the development of smarter validation methods, ensuring that the models learn in a meaningful way.

Substantial progress has been made in developing deep-learning models for clinical diagnosis. While excelling in diagnostics, the broader clinical decision-making process also involves establishing optimal follow-up intervals (TCU), crucial for prognosis and timely treatment. To fully support clinical practice, it is imperative that deep learning models contribute to both initial diagnosis and TCU prediction. However, relying on separate monolithic models is computationally demanding and lacks interpretability, hindering clinician trust. Our proposed bilateral model, emphasizing ophthalmological cases, offers both initial diagnoses and follow-up predictions, enhancing interpretability and trust in clinical applications as clinicians are more likely to trust recommendations, knowing the diagnosis used is correct. Inspired by clinical practice, the model integrates hierarchical inference and self-supervised learning techniques to enhance predictive accuracy and interpretability. Consisting of a sparse autoencoder, diagnosis classifier, and TCU classifier, the model leverages insights from clinicians and observations of ophthalmological datasets to capture salient features and facilitate robust learning. By employing shared weights for encoding and diagnosing each organ, the model optimizes efficiency and doubles the effective dataset size. Experimental results on an ophthalmological dataset demonstrate superior performance compared to baseline models, with the hierarchical inference structure providing valuable insights into the model's decision-making process. The bilateral model not only enhances predictive modeling for conditions affecting bilaterally symmetrical organs but also empowers clinicians with interpretable outputs crucial for informed clinical decision-making, thereby advancing clinical practice and improving patient care.

Analyzing patient health through irregular time series vital sign data demands innovative methods beyond conventional imputation techniques. This study introduces a novel approach diverging from prevailing attention-based models to explicitly capture temporal patient evolution. We adopt a paradigm where patients are viewed as dynamic systems evolving over time, with their vital signs encapsulating the system’s states. Our conceptual framework draws parallels to a Markov chain, exploring the transitions between states within a unit of time. To navigate the challenge of a vast state space, we employ a neural network to model expected transitions. Our method portrays the patient’s progression within one unit of time as the system evolves from one state to another, and forecasts states into the future. We outline the training process using irregular time series data and demonstrate its efficacy through analysis on two large vital sign data sets. Comparative analysis against attention-based models emphasizes the effectiveness and efficiency of our approach. This research heralds a promising avenue for patient vital sign analysis, providing insights into temporal patient evolution without relying on imputation methods, thereby enhancing predictive accuracy and interpretability of models.