The increasing complexity and workload of clinical radiology leads to inevitable oversights and mistakes in their use as diagnostic tools, causing delayed treatments and sometimes life-threatening harm to patients. While large language models (LLMs) have shown remarkable progress in many tasks, their utilities in detecting and correcting errors in radiology reporting are limited. This paper proposes a novel dual-knowledge infusion framework that enhances LLMs' capability for radiology report proofreading through systematic integration of medical expertise. Specifically, the knowledge infusion combines medical knowledge graph distillation (MKGD) with external knowledge retrieval (EXKR), enabling an effective automated approach in tackling mistakes in radiology reporting. By decomposing the complex proofreading task into three specialized stages of detection, localization, and correction, our method mirrors the systematic review process employed by expert radiologists, ensuring both precision and clinical interpretability. To perform a robust, clinically relevant evaluation, a comprehensive benchmark is also proposed using real-world radiology reports with real-world error patterns, including speech recognition confusions, terminology ambiguities, and template-related inconsistencies. Extensive evaluations across multiple LLM architectures demonstrate substantial improvements of our approach: up to 31.56% increase in error detection accuracy and 37.4% reduction in processing time. Human evaluation by radiologists confirms superior clinical relevance and factual consistency compared to existing approaches.

Imane Guellil, Salomé Andres, Atul Anand, Bruce Guthrie, Huayu Zhang, Abul Hasan, Honghan Wu, Beatrice Alex. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

OBJECTIVE: Due to heterogeneity and limited medical data in primary healthcare services (PHS), assessing the psychological risk of type 2 diabetes mellitus (T2DM) patients in PHS is difficult. Using unsupervised contrastive pre-training, we proposed a deep learning framework named depression and anxiety prediction (DAP) to predict depression and anxiety in T2DM patients. MATERIALS AND METHODS: The DAP model consists of two sub-models. Firstly, the pre-trained model of DAP used unlabeled discharge records of 85 085 T2DM patients from the First Affiliated Hospital of Nanjing Medical University for unsupervised contrastive learning on heterogeneous electronic health records (EHRs). Secondly, the fine-tuned model of DAP used case-control cohorts (17 491 patients) selected from 149 596 T2DM patients' EHRs in the Nanjing Health Information Platform (NHIP). The DAP model was validated in 1028 patients from PHS in NHIP. Evaluation included receiver operating characteristic area under the curve (ROC-AUC) and precision-recall area under the curve (PR-AUC), and decision curve analysis (DCA). RESULTS: The pre-training step allowed the DAP model to converge at a faster rate. The fine-tuned DAP model significantly outperformed the baseline models (logistic regression, extreme gradient boosting, and random forest) with ROC-AUC of 0.91±0.028 and PR-AUC of 0.80±0.067 in 10-fold internal validation, and with ROC-AUC of 0.75 ± 0.045 and PR-AUC of 0.47 ± 0.081 in external validation. The DCA indicate the clinical potential of the DAP model. CONCLUSION: The DAP model effectively predicted post-discharge depression and anxiety in T2DM patients from PHS, reducing data fragmentation and limitations. This study highlights the DAP model's potential for early detection and intervention in depression and anxiety, improving outcomes for diabetes patients.

MOTIVATION: Methods for concept recognition (CR) in clinical texts have largely been tested on abstracts or articles from the medical literature. However, texts from electronic health records (EHRs) frequently contain spelling errors, abbreviations, and other nonstandard ways of representing clinical concepts. RESULTS: Here, we present a method inspired by the BLAST algorithm for biosequence alignment that screens texts for potential matches on the basis of matching k-mer counts and scores candidates based on conformance to typical patterns of spelling errors derived from 2.9 million clinical notes. Our method, the Term-BLAST-like alignment tool (TBLAT) leverages a gold standard corpus for typographical errors to implement a sequence alignment-inspired method for efficient entity linkage. We present a comprehensive experimental comparison of TBLAT with five widely used tools. Experimental results show an increase of 10% in recall on scientific publications and 20% increase in recall on EHR records (when compared against the next best method), hence supporting a significant enhancement of the entity linking task. The method can be used stand-alone or as a complement to existing approaches. AVAILABILITY AND IMPLEMENTATION: Fenominal is a Java library that implements TBLAT for named CR of Human Phenotype Ontology terms and is available at https://github.com/monarch-initiative/fenominal under the GNU General Public License v3.0.

AI technologies are being increasingly tested and applied in critical environments including healthcare. Without an effective way to detect and mitigate AI induced inequalities, AI might do more harm than good, potentially leading to the widening of underlying inequalities. This paper proposes a generic allocation-deterioration framework for detecting and quantifying AI induced inequality. Specifically, AI induced inequalities are quantified as the area between two allocation-deterioration curves. To assess the framework’s performance, experiments were conducted on ten synthetic datasets (N>33,000) generated from HiRID - a real-world Intensive Care Unit (ICU) dataset, showing its ability to accurately detect and quantify inequality proportionally to controlled inequalities. Extensive analyses were carried out to quantify health inequalities (a) embedded in two real-world ICU datasets; (b) induced by AI models trained for two resource allocation scenarios. Results showed that compared to men, women had up to 33% poorer deterioration in markers of prognosis when admitted to HiRID ICUs. All four AI models assessed were shown to induce significant inequalities (2.45% to 43.2%) for non-White compared to White patients. The models exacerbated data embedded inequalities significantly in 3 out of 8 assessments, one of which was >9 times worse.

The ability to perform accurate prognosis is crucial for proactive clinical decision making, informed resource management and personalised care. Existing outcome prediction models suffer from a low recall of infrequent positive outcomes. We present a highly-scalable and robust machine learning framework to automatically predict adversity represented by mortality and ICU admission and readmission from time-series of vital signs and laboratory results obtained within the first 24 hours of hospital admission. The stacked ensemble platform comprises two components: a) an unsupervised LSTM Autoencoder that learns an optimal representation of the time-series, using it to differentiate the less frequent patterns which conclude with an adverse event from the majority patterns that do not, and b) a gradient boosting model, which relies on the constructed representation to refine prediction by incorporating static features. The model is used to assess a patient's risk of adversity and provides visual justifications of its prediction. Results of three case studies show that the model outperforms existing platforms in ICU and general ward settings, achieving average Precision-Recall Areas Under the Curve (PR-AUCs) of 0.891 (95% CI: 0.878-0.939) for mortality and 0.908 (95% CI: 0.870-0.935) in predicting ICU admission and readmission.

Network-based gene prioritization algorithms are designed to prioritize disease-associated genes based on known ones using biological networks of protein interactions, gene-disease associations (GDAs) and other relationships between biological entities. Various algorithms have been developed based on different mechanisms, but it is not obvious which algorithm is optimal for a specific disease. To address this issue, we benchmarked multiple algorithms for their application in cerebral small vessel disease (cSVD). We curated protein-gene interactions (PGIs) and GDAs from databases and assembled PGI networks and disease-gene heterogeneous networks. A screening of algorithms resulted in seven representative algorithms to be benchmarked. Performance of algorithms was assessed using both leave-one-out cross-validation (LOOCV) and external validation with MEGASTROKE genome-wide association study (GWAS). We found that random walk with restart on the heterogeneous network (RWRH) showed best LOOCV performance, with median LOOCV rediscovery rank of 185.5 (out of 19 463 genes). The GenePanda algorithm had most GWAS-confirmable genes in top 200 predictions, while RWRH had best ranks for small vessel stroke-associated genes confirmed in GWAS. In conclusion, RWRH has overall better performance for application in cSVD despite its susceptibility to bias caused by degree centrality. Choice of algorithms should be determined before applying to specific disease. Current pure network-based gene prioritization algorithms are unlikely to find novel disease-associated genes that are not associated with known ones. The tools for implementing and benchmarking algorithms have been made available and can be generalized for other diseases.

OBJECTIVE: Risk prediction models are widely used to inform evidence-based clinical decision making. However, few models developed from single cohorts can perform consistently well at population level where diverse prognoses exist (such as the SARS-CoV-2 [severe acute respiratory syndrome coronavirus 2] pandemic). This study aims at tackling this challenge by synergizing prediction models from the literature using ensemble learning. MATERIALS AND METHODS: In this study, we selected and reimplemented 7 prediction models for COVID-19 (coronavirus disease 2019) that were derived from diverse cohorts and used different implementation techniques. A novel ensemble learning framework was proposed to synergize them for realizing personalized predictions for individual patients. Four diverse international cohorts (2 from the United Kingdom and 2 from China; N = 5394) were used to validate all 8 models on discrimination, calibration, and clinical usefulness. RESULTS: Results showed that individual prediction models could perform well on some cohorts while poorly on others. Conversely, the ensemble model achieved the best performances consistently on all metrics quantifying discrimination, calibration, and clinical usefulness. Performance disparities were observed in cohorts from the 2 countries: all models achieved better performances on the China cohorts. DISCUSSION: When individual models were learned from complementary cohorts, the synergized model had the potential to achieve better performances than any individual model. Results indicate that blood parameters and physiological measurements might have better predictive powers when collected early, which remains to be confirmed by further studies. CONCLUSIONS: Combining a diverse set of individual prediction models, the ensemble method can synergize a robust and well-performing model by choosing the most competent ones for individual patients.

Many areas of research are characterised by the deluge of large-scale highly-dimensional time-series data. However, using the data available for prediction and decision making is hampered by the current lag in our ability to uncover and quantify true interactions that explain the outcomes. We are interested in areas such as intensive care medicine, which are characterised by i) continuous monitoring of multivariate variables and non-uniform sampling of data streams, ii) the outcomes are generally governed by interactions between a small set of rare events, iii) these interactions are not necessarily definable by specific values (or value ranges) of a given group of variables, but rather, by the deviations of these values from the normal state recorded over time, iv) the need to explain the predictions made by the model. Here, while numerous data mining models have been formulated for outcome prediction, they are unable to explain their predictions. We present a model for uncovering interactions with the highest likelihood of generating the outcomes seen from highly-dimensional time series data. Interactions among variables are represented by a relational graph structure, which relies on qualitative abstractions to overcome non-uniform sampling and to capture the semantics of the interactions corresponding to the changes and deviations from normality of variables of interest over time. Using the assumption that similar templates of small interactions are responsible for the outcomes (as prevalent in the medical domains), we reformulate the discovery task to retrieve the most-likely templates from the data. Experiments on sepsis prediction using real Intensive Care Unit (ICU) data demonstrates that the discovered interaction templates are semantically meaningful within the domain, and using them as features in a prediction task produces a superior performance than when using the raw values of the predictors.

OBJECTIVES: Current machine learning models aiming to predict sepsis from electronic health records (EHR) do not account 20 for the heterogeneity of the condition despite its emerging importance in prognosis and treatment. This work demonstrates the added value of stratifying the types of organ dysfunction observed in patients who develop sepsis in the intensive care unit (ICU) in improving the ability to recognize patients at risk of sepsis from their EHR data. MATERIALS AND METHODS: Using an ICU dataset of 13 728 records, we identify clinically significant sepsis subpopulations with distinct organ dysfunction patterns. We perform classification experiments with random forest, gradient boost trees, and support vector machines, using the identified subpopulations to distinguish patients who develop sepsis in the ICU from those who do not. RESULTS: The classification results show that features selected using sepsis subpopulations as background knowledge yield a superior performance in distinguishing septic from non-septic patients regardless of the classification model used. The improved performance is especially pronounced in specificity, which is a current bottleneck in sepsis prediction machine learning models. CONCLUSION: Our findings can steer machine learning efforts toward more personalized models for complex conditions including sepsis.

Objective: Unlocking the data contained within both structured and unstructured components of electronic health records (EHRs) has the potential to provide a step change in data available for secondary research use, generation of actionable medical insights, hospital management, and trial recruitment. To achieve this, we implemented SemEHR, an open source semantic search and analytics tool for EHRs. Methods: SemEHR implements a generic information extraction (IE) and retrieval infrastructure by identifying contextualized mentions of a wide range of biomedical concepts within EHRs. Natural language processing annotations are further assembled at the patient level and extended with EHR-specific knowledge to generate a timeline for each patient. The semantic data are serviced via ontology-based search and analytics interfaces. Results: SemEHR has been deployed at a number of UK hospitals, including the Clinical Record Interactive Search, an anonymized replica of the EHR of the UK South London and Maudsley National Health Service Foundation Trust, one of Europe's largest providers of mental health services. In 2 Clinical Record Interactive Search-based studies, SemEHR achieved 93% (hepatitis C) and 99% (HIV) F-measure results in identifying true positive patients. At King's College Hospital in London, as part of the CogStack program (github.com/cogstack), SemEHR is being used to recruit patients into the UK Department of Health 100 000 Genomes Project (genomicsengland.co.uk). The validation study suggests that the tool can validate previously recruited cases and is very fast at searching phenotypes; time for recruitment criteria checking was reduced from days to minutes. Validated on open intensive care EHR data, Medical Information Mart for Intensive Care III, the vital signs extracted by SemEHR can achieve around 97% accuracy. Conclusion: Results from the multiple case studies demonstrate SemEHR's efficiency: weeks or months of work can be done within hours or minutes in some cases. SemEHR provides a more comprehensive view of patients, bringing in more and unexpected insight compared to study-oriented bespoke IE systems. SemEHR is open source, available at https://github.com/CogStack/SemEHR.

Due to the increasing volume of and interconnections between semantic datasets, it becomes a challenging task for novice users to know what are included in a dataset, how they can make use of them, and particularly, what queries should be asked. In this paper we analyse several types of candidate insightful queries and propose a framework to generate such queries and identify their relations. To verify our approach, we implemented our framework and evaluated its performance with benchmark and real world datasets.