Rare and low-prevalence conditions in intensive care units present persistent challenges for clinicians and health systems. Patients affected by these conditions often experience higher mortality, longer stays and greater readmission risk than those with common diagnoses. Despite major advances in artificial intelligence for critical care, most predictive models are trained on large cohorts dominated by common conditions, leaving rare conditions underserved. Data scarcity, variation in clinical practice across institutions and substantial heterogeneity within each condition limit the reliability of existing approaches. Addressing these gaps requires methods that can learn from limited data without diluting clinical relevance. A recently developed deep learning framework responds to this need by combining generalisable learning from diverse electronic health records with targeted adaptation to clinically similar conditions, aiming to improve outcome prediction for rare conditions in the ICU.
Addressing Scarcity and Heterogeneity in ICU Data
Rare conditions in the ICU include both formally classified rare diseases and conditions that occur infrequently in critical care settings. These patients often face delayed diagnosis, limited specialist expertise and complex multisystem involvement, all of which contribute to poorer outcomes and higher resource use. Standard deep learning models struggle in this context because they require large, homogeneous datasets to achieve stable performance. Training exclusively on rare-condition data leads to weak generalisation, while pooling data from many conditions can introduce noise and degrade performance, particularly in multi-centre settings.
The proposed framework tackles this problem through a two-part strategy. First, it learns condition-agnostic representations by pre-training on time-series data from diverse ICU populations using self-supervised learning. This stage captures general temporal patterns in physiological data without focusing on any single diagnosis. Second, it addresses intra-condition heterogeneity by selectively adapting knowledge from clinically similar conditions rather than from all available data. Clinical similarity is identified using a condition knowledge graph that encodes relationships based on diagnosis co-occurrence, similarities in recorded variables and shared drug usage. This design allows the model to increase effective data volume while avoiding irrelevant or misleading information.
Knowledge-Guided Adaptation Across Clinical Tasks
The framework was evaluated using two widely used ICU datasets representing different care environments. One dataset reflects a single-centre setting with detailed longitudinal records, while the other captures multi-centre variability across more than 200 hospitals. Conditions were defined as rare when their prevalence was fewer than one case per 2,000 patients, and only those with at least one positive outcome sample were included. Five clinically relevant prediction tasks were assessed: 90-day mortality after discharge, 30-day readmission, ICU mortality, remaining length of stay and phenotyping.
Across all tasks and datasets, the framework consistently achieved higher precision–recall performance than baseline models designed for ICU outcome prediction and methods developed specifically for limited-data scenarios. Performance gains were observed for both short-term and post-discharge outcomes. Importantly, the approach also exceeded the performance of established ICU mortality scoring systems when predicting mortality for rare conditions in a multi-centre cohort. This suggests that targeted learning from clinically similar conditions can offer advantages over traditional scoring systems that rely on population-level averages dominated by common diagnoses.
Further analyses demonstrated that not all components contributed equally across tasks. Knowledge-guided domain selection emerged as a critical element, particularly for outcomes influenced by comorbidities and operational factors. Removing this component led to marked performance drops across datasets. Condition-agnostic pre-training supported generalisation, especially where training data were extremely limited, while joint adversarial domain adaptation helped align outcome distributions between rare and similar conditions. Together, these elements enabled robust performance across diverse clinical prediction tasks.
Must Read:Interpretable ML Improves Post-Transplant Risk Prediction
Adaptability, Generalisation and Clinical Relevance
A series of case studies explored how the framework adapts to different data environments and clinical scenarios. One analysis examined how many source conditions should be included during adaptation. Performance improved as more similar conditions were added, peaking when only a small proportion of the most similar conditions were selected. Including larger numbers of conditions introduced noise and reduced accuracy, particularly in multi-centre data. This finding challenges the assumption that more data always lead to better performance and highlights the value of selective learning.
Another analysis assessed sensitivity to the completeness of the condition knowledge graph. Optimal performance depended on dataset characteristics, with sparse graphs favouring multi-centre data and denser graphs benefiting single-centre data. The framework also demonstrated an ability to generalise to common conditions under artificial data scarcity. When trained on a fraction of available data for a common ICU diagnosis, it achieved comparable or better performance than standard models trained on full datasets for several outcomes. This suggests potential utility in real-world settings where access to comprehensive data is constrained.
The framework’s selection of source conditions was largely data-driven rather than aligned with hierarchical diagnostic codes. In several cases, clinically plausible similarities were identified across different diagnostic categories, reflecting shared physiological or treatment-related patterns rather than formal classifications. This behaviour enhances interpretability and supports clinical confidence in the model’s predictions.
By combining condition-agnostic learning with knowledge-guided adaptation, this deep learning framework offers a structured response to the dual challenges of data scarcity and heterogeneity in rare ICU conditions. Evaluations across multiple datasets and clinical tasks demonstrate consistent improvements over existing models and traditional scoring systems, particularly for outcomes where rare conditions are poorly served by current tools. The findings highlight the importance of selective data integration and clinically informed adaptation in predictive modelling. Such approaches point towards more equitable and accurate risk stratification for patients with rare conditions, with potential benefits for clinical decision-making, resource allocation and patient outcomes in intensive care settings.
Source: npj digital medicine
Image Credit: iStock
References:
Zhu M, Liu Y, Luo Z et al. (2026) Bridging data gaps of rare conditions in ICU: a multi-disease adaptation approach for clinical prediction. npj Digit Med; 9, 7.
