Search

Postoperative delirium affects 10–60% of elderly surgical patients and is linked to longer hospital stays, cognitive decline, and increased mortality. Although machine learning models have been developed to predict this condition using perioperative data, most rely on point predictions that fail to express uncertainty, limiting their clinical reliability in high-stakes surgical decision-making. These models often report a single risk estimate without indicating whether predictions are supported by strong or sparse evidence, which can lead to overconfidence and potential patient harm in vulnerable populations with heterogeneous frailty and comorbidity profiles. We argue that Bayesian deep learning is essential for postoperative delirium prediction because it provides distributional outputs and uncertainty estimates that allow clinicians to assess prediction reliability. Incorporating uncertainty quantification can transform these models from opaque tools into clinically trustworthy decision aids. We recommend that uncertainty reporting be required in all predictive models for postoperative delirium and that regulatory and publication standards enforce the use of Bayesian approaches. Overall, replacing point estimates with distributional predictions is necessary to improve safety and clinical utility in perioperative care of elderly patients.

Acute ischemic stroke prediction from electronic health record time series data holds significant potential for enabling early intervention and reducing long-term disability. LSTMs have been widely used to model clinical sequences such as vital signs and laboratory trends, showing strong performance in stroke-related prediction tasks from 2018–2022. However, their sequential nature limits scalability and long-range dependency modeling in large EHR datasets. Transformers, despite transforming sequence modeling in other domains since 2017, remain underused in stroke prediction compared to LSTMs. Although early healthcare studies suggest potential benefits of attention-based models, robust validation in acute ischemic stroke contexts is still limited. Transformers offer advantages in parallel processing, long-range dependency modeling, and interpretability, but require more data and computational resources. They are likely to complement rather than replace LSTMs, with hybrid architectures providing a balanced solution for clinical time series analysis. Key themes include long-range dependency capture, parallel computation, interpretability, and data efficiency trade-offs between LSTMs and transformers. Hybrid LSTM–transformer models may offer improved performance and practicality for stroke prediction, with model selection depending on data scale and clinical constraints. Further benchmarking is needed to determine when transformers or hybrid models outperform LSTMs, guiding the development of more effective stroke prediction systems.

Pancreatic cancer is highly lethal, and surgical resection is the only curative option. Preoperative assessment using contrast-enhanced CT is essential for determining tumor resectability based on involvement of key vessels such as the superior mesenteric artery, celiac trunk, and portal vein. Accurate pancreatic tumor segmentation is difficult due to unclear boundaries, low contrast with surrounding tissue, and proximity to major vessels. Manual segmentation is slow, subjective, and inconsistent, especially in borderline cases, while tumor-associated fibrosis further obscures lesion margins. We propose a deep learning-based framework using an attention-enhanced U-Net with multi-scale feature fusion and deep supervision for tumor segmentation and resectability assessment. The model incorporates attention gates, atrous spatial pyramid pooling, and auxiliary losses at multiple decoder levels to improve feature learning and gradient flow. A pre-trained encoder extracts hierarchical features refined by attention mechanisms in skip connections. A multi-scale decoder reconstructs segmentation maps, supported by deep supervision at different resolutions. A parallel branch models tumor–vessel spatial relationships using distance maps to improve resectability classification. This framework enables automated pancreatic tumor segmentation and resectability evaluation from CT scans, improving accuracy, interpretability, and clinical utility. Validation on datasets such as Pancreas-CT and Medical Segmentation Decathlon is recommended.

Rare pediatric tumors like sarcomas, neuroblastoma, medulloblastoma, and retinoblastoma pose a challenge for developing deep learning models due to the limited availability of histopathology images, which are distributed across multiple institutions. This scarcity is compounded by privacy concerns, as whole-slide images often contain sensitive clinical and genomic data, and generative adversarial networks (GANs) risk memorizing and leaking training samples. To address this, a differentially private GAN framework is proposed for synthesizing high-resolution histopathology patches of rare pediatric cancers. The framework incorporates a generator for image synthesis, a discriminator for realism assessment, per-sample gradient clipping, Gaussian noise injection, and a privacy accountant, ensuring provable privacy guarantees during the training process. The synthetic images generated can aid in data augmentation, model pre-training, and benchmarking without exposing identifiable pathology data, offering a privacy-preserving solution for dataset augmentation while emphasizing the importance of clinical validation.

Anticoagulation management requires balancing multiple factors such as bleeding risk, thromboembolic risk, drug interactions, and renal function. Deep learning can assist in risk prediction, but its effectiveness relies on clinicians' ability to understand and verify the recommendations. Black-box models may recommend actions without providing clear explanations. In contrast, clinical guidelines are rule-based but not directly executable by neural models. This article introduces a neuro-symbolic XAI framework that combines deep learning predictions with explicit clinical guidelines. It includes a neural prediction module, a symbolic reasoning engine, and an integration layer for traceable justifications. The neuro-symbolic approach connects data-driven predictions to clinical rules, improving auditability and trustworthiness in decision support. This framework aims to enhance anticoagulation management by providing verifiable, clinician-understandable decision support, focusing on explainability-by-design.