OBJECTIVE: To optimise the identification of patients presenting with pain in emergency department (ED) settings with limited resources using multiple transfer learning techniques. METHODS: Two strategies were explored: (1) fine-tuning a pre-trained language model, previously fine-tuned on data from a well-resourced ED, using labelled data from a target ED, and (2) continual pre-training using task-specific unlabelled data to enhance clinical text classification. RESULTS: With 2000 labelled samples from a target ED, the combined strategies achieved an F1-score of 92%, demonstrating significant benefits of transfer learning in resource-constrained settings. DISCUSSION: Accurately identifying pain in patients upon arrival to the ED is crucial for timely and effective treatment. Findings suggest that combining both transfer learning strategies can significantly enhance pain identification performances in resource-constrained settings. CONCLUSION: Combining fine-tuning on labelled data and continual pre-training on unlabelled data has potential to optimise model performance in both resource-constrained and well-resourced settings, highlighting the broader applicability and potential of these techniques for improving clinical text classification.

Large language models (LLMs) demonstrate impressive capabilities in generating human-like content and have much potential to improve the performance and efficiency of healthcare. An important application of LLMs is to generate synthetic clinical reports that could alleviate the burden of annotating and collecting real-world data in training AI models. Meanwhile, there could be concerns and limitations in using commercial LLMs to handle sensitive clinical data. In this study, we examined the use of open-source LLMs as an alternative to generate synthetic radiology reports to supplement real-world annotated data. We found LLMs hosted locally can achieve similar performance compared to ChatGPT and GPT-4 in augmenting training data for the downstream report classification task of identifying misdiagnosed fractures. We also examined the predictive value of using synthetic reports alone for training downstream models, where our best setting achieved more than 90 % of the performance using real-world data. Overall, our findings show that open-source, local LLMs can be a favourable option for creating synthetic clinical reports for downstream tasks. • Popular open-source LLMs were compared with proprietary LLMs in generating synthetic radiology reports. • Synthetic reports were applied to train downstream report classification models to identify misdiagnosed limb fractures. • Open-source LLMs can achieve similar performances as proprietary LLMs in generating reports with high predictive value. • Impact of reports from different hospital sources were investigated and directions for further research were identified.

Clinical predictive analysis is a crucial task with numerous applications and has been extensively studied using machine learning approaches. Clinical notes, a vital data source, have been employed to develop natural language processing (NLP) models for risk prediction in healthcare with robust performance. However, clinical notes vary considerably in text composition-written by diverse healthcare providers for different purposes-and the impact of these variations on NLP modeling is also underexplored. It also remains uncertain whether the recent Large Language Models (LLMs) with instruction-following capabilities can effectively handle the risk prediction task out-of-the-box, especially when using routinely collected clinical notes instead of polished text. We address these two important research questions in the context of in-hospital mortality prediction within the critical care setting. Specifically, we propose a supervised hierarchical network with note-specific modules to account for variations across different note categories, and provide a detailed comparison with strong supervised baselines and LLMs. We benchmark 34 instruction-following LLMs based on zero-shot, few-shot, and chain-of-thought prompting with diverse prompt templates. Our results demonstrate that the note-specific network delivers improved risk prediction performance compared to established supervised baselines from both measurement-based and text-based modeling. In contrast, LLMs consistently underperform on this critical task, despite their remarkable performances in other domains. This highlights important limitations and raises caution regarding the use of LLMs for risk assessment in the critical setting. Additionally, we show that the proposed model can be leveraged to select informative clinical notes to enhance the training of other models.

Encouraged by the success of pretrained Transformer models in many natural language processing tasks, their use for International Classification of Diseases (ICD) coding tasks is now actively being explored. In this study, we investigated two existing Transformer-based models (PLM-ICD and XR-Transformer) and proposed a novel Transformer-based model (XR-LAT), aiming to address the extreme label set and long text classification challenges that are posed by automated ICD coding tasks. The Transformer-based model PLM-ICD, which currently holds the state-of-the-art (SOTA) performance on the ICD coding benchmark datasets MIMIC-III and MIMIC-II, was selected as our baseline model for further optimisation on both datasets. In addition, we extended the capabilities of the leading model in the general extreme multi-label text classification domain, XR-Transformer, to support longer sequences and trained it on both datasets. Moreover, we proposed a novel model, XR-LAT, which was also trained on both datasets. XR-LAT is a recursively trained model chain on a predefined hierarchical code tree with label-wise attention, knowledge transferring and dynamic negative sampling mechanisms. Our optimised PLM-ICD models, which were trained with longer total and chunk sequence lengths, significantly outperformed the current SOTA PLM-ICD models, and achieved the highest micro-F1 scores of 60.8 % and 50.9 % on MIMIC-III and MIMIC-II, respectively. The XR-Transformer model, although SOTA in the general domain, did not perform well across all metrics. The best XR-LAT based models obtained results that were competitive with the current SOTA PLM-ICD models, including improving the macro-AUC by 2.1 % and 5.1 % on MIMIC-III and MIMIC-II, respectively. Our optimised PLM-ICD models are the new SOTA models for automated ICD coding on both datasets, while our novel XR-LAT models perform competitively with the previous SOTA PLM-ICD models.

OBJECTIVE: With the increasing amount and growing variety of healthcare data, multimodal machine learning supporting integrated modeling of structured and unstructured data is an increasingly important tool for clinical machine learning tasks. However, it is non-trivial to manage the differences in dimensionality, volume, and temporal characteristics of data modalities in the context of a shared target task. Furthermore, patients can have substantial variations in the availability of data, while existing multimodal modeling methods typically assume data completeness and lack a mechanism to handle missing modalities. METHODS: We propose a Transformer-based fusion model with modality-specific tokens that summarize the corresponding modalities to achieve effective cross-modal interaction accommodating missing modalities in the clinical context. The model is further refined by inter-modal, inter-sample contrastive learning to improve the representations for better predictive performance. We denote the model as Attention-based cRoss-MOdal fUsion with contRast (ARMOUR). We evaluate ARMOUR using two input modalities (structured measurements and unstructured text), six clinical prediction tasks, and two evaluation regimes, either including or excluding samples with missing modalities. RESULTS: Our model shows improved performances over unimodal or multimodal baselines in both evaluation regimes, including or excluding patients with missing modalities in the input. The contrastive learning improves the representation power and is shown to be essential for better results. The simple setup of modality-specific tokens enables ARMOUR to handle patients with missing modalities and allows comparison with existing unimodal benchmark results. CONCLUSION: We propose a multimodal model for robust clinical prediction to achieve improved performance while accommodating patients with missing modalities. This work could inspire future research to study the effective incorporation of multiple, more complex modalities of clinical data into a single model.

International Classification of Diseases (ICD) coding plays an important role in systematically classifying morbidity and mortality data. In this study, we propose a hierarchical label-wise attention Transformer model (HiLAT) for the explainable prediction of ICD codes from clinical documents. HiLAT firstly fine-tunes a pretrained Transformer model to represent the tokens of clinical documents. We subsequently employ a two-level hierarchical label-wise attention mechanism that creates label-specific document representations. These representations are in turn used by a feed-forward neural network to predict whether a specific ICD code is assigned to the input clinical document of interest. We evaluate HiLAT using hospital discharge summaries and their corresponding ICD-9 codes from the MIMIC-III database. To investigate the performance of different types of Transformer models, we develop ClinicalplusXLNet, which conducts continual pretraining from XLNet-Base using all the MIMIC-III clinical notes. The experiment results show that the F1 scores of the HiLAT + ClinicalplusXLNet outperform the previous state-of-the-art models for the top-50 most frequent ICD-9 codes from MIMIC-III. Visualisations of attention weights present a potential explainability tool for checking the face validity of ICD code predictions.

Electronic Medical Records (EMRs) contain clinical narrative text that is of great potential value to medical researchers. However, this information is mixed with Personally Identifiable Information (PII) that presents risks to patient and clinician confidentiality. This paper presents an end-to-end de-identification framework to automatically remove PII from Australian hospital discharge summaries. Our corpus included 600 hospital discharge summaries which were extracted from the EMRs of two principal referral hospitals in Sydney, Australia. Our end-to-end de-identification framework consists of three components: (1) Annotation: labelling of PII in the 600 hospital discharge summaries using five pre-defined categories: person, address, date of birth, individual identification number, phone/fax number; (2) Modelling: training six named entity recognition (NER) deep learning base-models on balanced and imbalanced datasets; and evaluating ensembles that combine all six base-models, the three base-models with the best F1 scores and the three base-models with the best recall scores respectively, using token-level majority voting and stacking methods; and (3) De-identification: removing PII from the hospital discharge summaries. Our results showed that the ensemble model combined using the stacking Support Vector Machine (SVM) method on the three base-models with the best F1 scores achieved excellent results with a F1 score of 99.16% on the test set of our corpus. We also evaluated the robustness of our modelling component on the 2014 i2b2 de-identification dataset. Our ensemble model, which uses the token-level majority voting method on all six base-models, achieved the highest F1 score of 96.24% at strict entity matching and the highest F1 score of 98.64% at binary token-level matching compared to two state-of-the-art methods. The end-to-end framework provides a robust solution to de-identifying clinical narrative corpuses safely. It can easily be applied to any kind of clinical narrative documents.

One unintended consequence of the Electronic Health Records (EHR) implementation is the overuse of content-importing technology, such as copy-and-paste, that creates "bloated" notes containing large amounts of textual redundancy. Despite the rising interest in applying machine learning models to learn from real-patient data, it is unclear how the phenomenon of note bloat might affect the Natural Language Processing (NLP) models derived from these notes. Therefore, in this work we examine the impact of redundancy on deep learning-based NLP models, considering four clinical prediction tasks using a publicly available EHR database. We applied two deduplication methods to the hospital notes, identifying large quantities of redundancy, and found that removing the redundancy usually has little negative impact on downstream performances, and can in certain circumstances assist models to achieve significantly better results. We also showed it is possible to attack model predictions by simply adding note duplicates, causing changes of correct predictions made by trained models into wrong predictions. In conclusion, we demonstrated that EHR text redundancy substantively affects NLP models for clinical prediction tasks, showing that the awareness of clinical contexts and robust modeling methods are important to create effective and reliable NLP systems in healthcare contexts.

ICD coding is a process of assigning the International Classification of Disease diagnosis codes to clinical/medical notes documented by health professionals (e.g. clinicians). This process requires significant human resources, and thus is costly and prone to error. To handle the problem, machine learning has been utilized for automatic ICD coding. Previous state-of-the-art models were based on convolutional neural networks, using a single/several fixed window sizes. However, the lengths and interdependence between text fragments related to ICD codes in clinical text vary significantly, leading to the difficulty of deciding what the best window sizes are. In this paper, we propose a new label attention model for automatic ICD coding, which can handle both the various lengths and the interdependence of the ICD code related text fragments. Furthermore, as the majority of ICD codes are not frequently used, leading to the extremely imbalanced data issue, we additionally propose a hierarchical joint learning mechanism extending our label attention model to handle the issue, using the hierarchical relationships among the codes. Our label attention model achieves new state-of-the-art results on three benchmark MIMIC datasets, and the joint learning mechanism helps improve the performances for infrequent codes.

This article demonstrates the benefits of using sequence representations based on word embeddings to inform the seed selection and sample selection processes in an active learning pipeline for clinical information extraction. Seed selection refers to choosing an initial sample set to label to form an initial learning model. Sample selection refers to selecting informative samples to update the model at each iteration of the active learning process. Compared to supervised machine learning approaches, active learning offers the opportunity to build statistical classifiers with a reduced amount of training samples that require manual annotation. Reducing the manual annotation effort can support automating the clinical information extraction process. This is particularly beneficial in the clinical domain, where manual annotation is a time‐consuming and costly task, as it requires extensive labor from clinical experts. Our empirical findings demonstrate that (a) using sequence representations along with the length of sequence for seed selection shows potential towards more effective initial models, and (b) using sequence representations for sample selection leads to significantly lower manual annotation efforts, with up to 3% and 6% fewer tokens and concepts requiring annotation, respectively, compared to state‐of‐the‐art query strategies.

OBJECTIVE: This paper presents an automatic, active learning-based system for the extraction of medical concepts from clinical free-text reports. Specifically, (1) the contribution of active learning in reducing the annotation effort and (2) the robustness of incremental active learning framework across different selection criteria and data sets are determined. MATERIALS AND METHODS: The comparative performance of an active learning framework and a fully supervised approach were investigated to study how active learning reduces the annotation effort while achieving the same effectiveness as a supervised approach. Conditional random fields as the supervised method, and least confidence and information density as 2 selection criteria for active learning framework were used. The effect of incremental learning vs standard learning on the robustness of the models within the active learning framework with different selection criteria was also investigated. The following 2 clinical data sets were used for evaluation: the Informatics for Integrating Biology and the Bedside/Veteran Affairs (i2b2/VA) 2010 natural language processing challenge and the Shared Annotated Resources/Conference and Labs of the Evaluation Forum (ShARe/CLEF) 2013 eHealth Evaluation Lab. RESULTS: The annotation effort saved by active learning to achieve the same effectiveness as supervised learning is up to 77%, 57%, and 46% of the total number of sequences, tokens, and concepts, respectively. Compared with the random sampling baseline, the saving is at least doubled. CONCLUSION: Incremental active learning is a promising approach for building effective and robust medical concept extraction models while significantly reducing the burden of manual annotation.

This paper presents a new active learning query strategy for information extraction, called Domain Knowledge Informativeness (DKI). Active learning is often used to reduce the amount of annotation effort required to obtain training data for machine learning algorithms. A key component of an active learning approach is the query strategy, which is used to iteratively select samples for annotation. Knowledge resources have been used in information extraction as a means to derive additional features for sample representation. DKI is, however, the first query strategy that exploits such resources to inform sample selection. To evaluate the merits of DKI, in particular with respect to the reduction in annotation effort that the new query strategy allows to achieve, we conduct a comprehensive empirical comparison of active learning query strategies for information extraction within the clinical domain. The clinical domain was chosen for this work because of the availability of extensive structured knowledge resources which have often been exploited for feature generation. In addition, the clinical domain offers a compelling use case for active learning because of the necessary high costs and hurdles associated with obtaining annotations in this domain. Our experimental findings demonstrated that (1) amongst existing query strategies, the ones based on the classification model's confidence are a better choice for clinical data as they perform equally well with a much lighter computational load, and (2) significant reductions in annotation effort are achievable by exploiting knowledge resources within active learning query strategies, with up to 14% less tokens and concepts to manually annotate than with state-of-the-art query strategies.

OBJECTIVE: To classify automatically lung tumor-node-metastases (TNM) cancer stages from free-text pathology reports using symbolic rule-based classification. DESIGN: By exploiting report substructure and the symbolic manipulation of systematized nomenclature of medicine-clinical terms (SNOMED CT) concepts in reports, statements in free text can be evaluated for relevance against factors relating to the staging guidelines. Post-coordinated SNOMED CT expressions based on templates were defined and populated by concepts in reports, and tested for subsumption by staging factors. The subsumption results were used to build logic according to the staging guidelines to calculate the TNM stage. MEASUREMENTS: The accuracy measure and confusion matrices were used to evaluate the TNM stages classified by the symbolic rule-based system. The system was evaluated against a database of multidisciplinary team staging decisions and a machine learning-based text classification system using support vector machines. RESULTS: Overall accuracy on a corpus of pathology reports for 718 lung cancer patients against a database of pathological TNM staging decisions were 72%, 78%, and 94% for T, N, and M staging, respectively. The system's performance was also comparable to support vector machine classification approaches. CONCLUSION: A system to classify lung TNM stages from free-text pathology reports was developed, and it was verified that the symbolic rule-based approach using SNOMED CT can be used for the extraction of key lung cancer characteristics from free-text reports. Future work will investigate the applicability of using the proposed methodology for extracting other cancer characteristics and types.

Cancer staging provides a basis for planning clinical management, but also allows for meaningful analysis of cancer outcomes and evaluation of cancer care services. Despite this, stage data in cancer registries is often incomplete, inaccurate, or simply not collected. This article describes a prototype software system (Cancer Stage Interpretation System, CSIS) that automatically extracts cancer staging information from medical reports. The system uses text classification techniques to train support vector machines (SVMs) to extract elements of stage listed in cancer staging guidelines. When processing new reports, CSIS identifies sentences relevant to the staging decision, and subsequently assigns the most likely stage. The system was developed using a database of staging data and pathology reports for 710 lung cancer patients, then validated in an independent set of 179 patients against pathologic stage assigned by two independent pathologists. CSIS achieved overall accuracy of 74% for tumor (T) staging and 87% for node (N) staging, and errors were observed to mirror disagreements between human experts.

The use of visual attention (VA) spatial and temporal characteristics, monitored by a gaze-tracking device, to generate a region of interest (ROI) 'importance' map is proposed. A K-means clustering approach is adopted to group gaze location points into a number of clusters to represent the loci of regions of VA (or ROIs). Several metrics are then derived from the gaze positions and sequences to quantify the relative importance of the K-means clusters. An entropy-weighting strategy is adopted for the combination of these metrics to generate the ROI map. Results show that the ROI map is robust to the number of clusters and different gaze patterns, and can be used in progressive image coding/decoding to enhance the image quality in regions of interest.

The paper proposes a general framework for the importance coding of still images to maximise the interpretability versus bitrate performance. The interpretability of an image to achieve maximum content recognition is important in a diverse range of applications such as surveillance and medical. Importance coding aims to address this problem by prioritisation of the encoded image bit-stream based on the importance of regions in an image. Consequently, the most important features required for interpretability are encoded and transmitted earlier in the encoded image bit-stream. The notion of importance maps, which provide a systematic approach for the assignment of relative importance of regions in an image, are presented and its use in importance coding is developed. One highly desirable advantage of the proposed importance coding framework is that it can be implemented by the EBCOT (embedded block coding with optimised truncation) coder, which has been selected as the core-coding algorithm in JPEG 2000.