Delirium and incident dementia in hospital patients in New South Wales, Australia: retrospective cohort study.

OBJECTIVES: To determine the strength and nature of the association between delirium and incident dementia in a population of older adult patients without dementia at baseline. DESIGN: Retrospective cohort study using large scale hospital administrative data. SETTING: Public and private hospitals in New South Wales, Australia between July 2001 and March 2020. PARTICIPANTS: Data were extracted for 650 590 hospital patients aged ≥65 years. Diagnoses of dementia and delirium were identified from ICD-10 (international classification of diseases, 10th revision) codes. Patients with dementia at baseline were excluded. Delirium-no delirium pairs were identified by matching personal and clinical characteristics, and were followed for more than five years. MAIN OUTCOME MEASURES: Cox proportional hazards models and Fine-Gray hazard models were used to estimate the associations of delirium with death and incident dementia, respectively. Delirium-outcome dose-response associations were quantified, all analyses were performed in men and women separately, and sensitivity analyses were conducted. RESULTS: The study included 55 211 matched pairs (48% men, mean age 83.4 years, standard deviation 6.5 years). Collectively, 58% (n=63 929) of patients died and 17% (n=19 117) had a newly reported dementia diagnosis during 5.25 years of follow-up. Patients with delirium had 39% higher risk of death (hazard ratio 1.39, 95% confidence interval 1.37 to 1.41) and three times higher risk of incident dementia (subdistribution hazard ratio 3.00, 95% confidence interval 2.91 to 3.10) than patients without delirium. The association with dementia was stronger in men (P=0.004). Each additional episode of delirium was associated with a 20% increased risk of dementia (subdistribution hazard ratio 1.20, 95% confidence interval 1.18 to 1.23). CONCLUSIONS: The study findings suggest delirium was a strong risk factor for death and incident dementia among older adult patients. The data support a causal interpretation of the association between delirium and dementia. The clinical implications of delirium as a potentially modifiable risk factor for dementia are substantial.

How Well Do AI-Enabled Decision Support Systems Perform in Clinical Settings?

Real-world performance of machine learning (ML) models is crucial for safely and effectively embedding them into clinical decision support (CDS) systems. We examined evidence about the performance of contemporary ML-based CDS in clinical settings. A systematic search of four bibliographic databases identified 32 studies over a 5-year period. The CDS task, ML type, ML method and real-world performance was extracted and analysed. Most ML-based CDS supported image recognition and interpretation (n=12; 38%) and risk assessment (n=9; 28%). The majority used supervised learning (n=28; 88%) to train random forests (n=7; 22%) and convolutional neural networks (n=7; 22%). Only 12 studies reported real-world performance using heterogenous metrics; and performance degraded in clinical settings compared to model validation. The reporting of model performance is fundamental to ensuring safe and effective use of ML-based CDS in clinical settings. There remain opportunities to improve reporting.

Denoising Cough Sound Recordings Using Neural Networks.

Objective cough sound evaluation is useful in the diagnosis and management of respiratory diseases. However, the performance of cough sound analysis models can degrade in the presence of background noises common in everyday environments. This brings forward the need for cough sound denoising. This work utilizes a method for denoising cough sound recordings using signal processing and machine learning techniques, inspired by research in the field of speech enhancement. It uses supervised learning to find a mapping between the noisy and clean spectra of cough sound signals using a fully connected feed-forward neural network. The method is validated on a dataset of 300 manually annotated cough sound recordings corrupted with babble noise. The effect of various signal processing and neural network parameters on denoising performance is investigated. The method is shown to improve cough sound quality and intelligibility and outperform conventional denoising methods.

Domain-Specific Pre-training Improves Confidence in Whole Slide Image Classification.

Whole Slide Images (WSIs) or histopathology images are used in digital pathology. WSIs pose great challenges to deep learning models for clinical diagnosis, owing to their size and lack of pixel-level annotations. With the recent advancements in computational pathology, newer multiple-instance learning-based models have been proposed. Multiple-instance learning for WSIs necessitates creating patches and uses the encoding of these patches for diagnosis. These models use generic pre-trained models (ResNet-50 pre-trained on ImageNet) for patch encoding. The recently proposed KimiaNet, a DenseNet121 model pre-trained on TCGA slides, is a domain-specific pre-trained model. This paper shows the effect of domain-specific pre-training on WSI classification. To investigate the effect of domain-specific pre-training, we considered the current state-of-the-art multiple-instance learning models, 1) CLAM, an attention-based model, and 2) TransMIL, a self-attention-based model, and evaluated the models' confidence and predictive performance in detecting primary brain tumors - gliomas. Domain-specific pre-training improves the confidence of the models and also achieves a new state-of-the-art performance of WSI-based glioma subtype classification, showing a high clinical applicability in assisting glioma diagnosis. We will publicly share our code and experimental results at https://github.com/soham-chitnis10/WSI-domain-specific.

Effects of machine learning-based clinical decision support systems on decision-making, care delivery, and patient outcomes: a scoping review.

OBJECTIVE: This study aims to summarize the research literature evaluating machine learning (ML)-based clinical decision support (CDS) systems in healthcare settings. MATERIALS AND METHODS: We conducted a review in accordance with the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta Analyses extension for Scoping Review). Four databases, including PubMed, Medline, Embase, and Scopus were searched for studies published from January 2016 to April 2021 evaluating the use of ML-based CDS in clinical settings. We extracted the study design, care setting, clinical task, CDS task, and ML method. The level of CDS autonomy was examined using a previously published 3-level classification based on the division of clinical tasks between the clinician and CDS; effects on decision-making, care delivery, and patient outcomes were summarized. RESULTS: Thirty-two studies evaluating the use of ML-based CDS in clinical settings were identified. All were undertaken in developed countries and largely in secondary and tertiary care settings. The most common clinical tasks supported by ML-based CDS were image recognition and interpretation (n = 12) and risk assessment (n = 9). The majority of studies examined assistive CDS (n = 23) which required clinicians to confirm or approve CDS recommendations for risk assessment in sepsis and for interpreting cancerous lesions in colonoscopy. Effects on decision-making, care delivery, and patient outcomes were mixed. CONCLUSION: ML-based CDS are being evaluated in many clinical areas. There remain many opportunities to apply and evaluate effects of ML-based CDS on decision-making, care delivery, and patient outcomes, particularly in resource-constrained settings.

Clinical named entity recognition and relation extraction using natural language processing of medical free text: A systematic review.

BACKGROUND: Natural Language Processing (NLP) applications have developed over the past years in various fields including its application to clinical free text for named entity recognition and relation extraction. However, there has been rapid developments the last few years that there's currently no overview of it. Moreover, it is unclear how these models and tools have been translated into clinical practice. We aim to synthesize and review these developments. METHODS: We reviewed literature from 2010 to date, searching PubMed, Scopus, the Association of Computational Linguistics (ACL), and Association of Computer Machinery (ACM) libraries for studies of NLP systems performing general-purpose (i.e., not disease- or treatment-specific) information extraction and relation extraction tasks in unstructured clinical text (e.g., discharge summaries). RESULTS: We included in the review 94 studies with 30 studies published in the last three years. Machine learning methods were used in 68 studies, rule-based in 5 studies, and both in 22 studies. 63 studies focused on Named Entity Recognition, 13 on Relation Extraction and 18 performed both. The most frequently extracted entities were "problem", "test" and "treatment". 72 studies used public datasets and 22 studies used proprietary datasets alone. Only 14 studies defined clearly a clinical or information task to be addressed by the system and just three studies reported its use outside the experimental setting. Only 7 studies shared a pre-trained model and only 8 an available software tool. DISCUSSION: Machine learning-based methods have dominated the NLP field on information extraction tasks. More recently, Transformer-based language models are taking the lead and showing the strongest performance. However, these developments are mostly based on a few datasets and generic annotations, with very few real-world use cases. This may raise questions about the generalizability of findings, translation into practice and highlights the need for robust clinical evaluation.

Heart and brain traumatic stress biomarker analysis with and without machine learning: A scoping review.

The enigma of post-traumatic stress disorder (PTSD) is embedded in a complex array of physiological responses to stressful situations that result in disruptions in arousal and cognitions that characterise the psychological disorder. Deciphering these physiological patterns is complex, which has seen the use of machine learning (ML) grow in popularity. However, it is unclear to what extent ML has been used with physiological data, specifically, the electroencephalogram (EEG) and electrocardiogram (ECG) to further understand the physiological responses associated with PTSD. To better understand the use of EEG and ECG biomarkers, with and without ML, a scoping review was undertaken. A total of 124 papers based on adult samples were identified comprising 19 ML studies involving EEG and ECG. A further 21 studies using EEG data, and 84 studies employing ECG meeting all other criteria but not employing ML were included for comparison. Identified studies indicate classical ML methodologies currently dominate EEG and ECG biomarkers research, with derived biomarkers holding clinically relevant diagnostic implications for PTSD. Discussion of the emerging trends, algorithms used and their success is provided, along with areas for future research.

Reliability of crowdsourced data and patient-reported outcome measures in cough-based COVID-19 screening.

Mass community testing is a critical means for monitoring the spread of the COVID-19 pandemic. Polymerase chain reaction (PCR) is the gold standard for detecting the causative coronavirus 2 (SARS-CoV-2) but the test is invasive, test centers may not be readily available, and the wait for laboratory results can take several days. Various machine learning based alternatives to PCR screening for SARS-CoV-2 have been proposed, including cough sound analysis. Cough classification models appear to be a robust means to predict infective status, but collecting reliable PCR confirmed data for their development is challenging and recent work using unverified crowdsourced data is seen as a viable alternative. In this study, we report experiments that assess cough classification models trained (i) using data from PCR-confirmed COVID subjects and (ii) using data of individuals self-reporting their infective status. We compare performance using PCR-confirmed data. Models trained on PCR-confirmed data perform better than those trained on patient-reported data. Models using PCR-confirmed data also exploit more stable predictive features and converge faster. Crowd-sourced cough data is less reliable than PCR-confirmed data for developing predictive models for COVID-19, and raises concerns about the utility of patient reported outcome data in developing other clinical predictive models when better gold-standard data are available.

Colour adaptive generative networks for stain normalisation of histopathology images.

Deep learning has shown its effectiveness in histopathology image analysis, such as pathology detection and classification. However, stain colour variation in Hematoxylin and Eosin (H&E) stained histopathology images poses challenges in effectively training deep learning-based algorithms. To alleviate this problem, stain normalisation methods have been proposed, with most of the recent methods utilising generative adversarial networks (GAN). However, these methods are either trained fully with paired images from the target domain (supervised) or with unpaired images (unsupervised), suffering from either large discrepancy between domains or risks of undertrained/overfitted models when only the target domain images are used for training. In this paper, we introduce a colour adaptive generative network (CAGAN) for stain normalisation which combines both supervised learning from target domain and unsupervised learning from source domain. Specifically, we propose a dual-decoder generator and force consistency between their outputs thus introducing extra supervision which benefits from extra training with source domain images. Moreover, our model is immutable to stain colour variations due to the use of stain colour augmentation. We further implement histogram loss to ensure the processed images are coloured with the target domain colours regardless of their content differences. Extensive experiments on four public histopathology image datasets including TCGA-IDH, CAMELYON16, CAMELYON17 and BreakHis demonstrate that our proposed method produces high quality stain normalised images which improve the performance of benchmark algorithms by 5% to 10% compared to baselines not using normalisation.

Comparing health outcomes between coronary interventions in frail patients aged 75 years or older with acute coronary syndrome: a systematic review.

PURPOSE: To assess current evidence comparing the impact of available coronary interventions in frail patients aged 75 years or older with different subtypes of acute coronary syndrome (ACS) on health outcomes. METHODS: Scopus, Embase and PubMed were systematically searched in May 2022 for studies comparing outcomes between coronary interventions in frail older patients with ACS. Studies were excluded if they provided no objective assessment of frailty during the index admission, under-represented patients aged 75 years or older, or included patients with non-ACS coronary disease without presenting results for the ACS subgroup. Following data extraction from the included studies, a qualitative synthesis of results was undertaken. RESULTS: Nine studies met all eligibility criteria. All eligible studies were observational. Substantial heterogeneity was observed across study designs regarding ACS subtypes included, frailty assessments used, coronary interventions compared, and outcomes studied. All studies were assessed to be at high risk of bias. Notably, adjustment for confounders was limited or not adequately reported in all studies. The comparative assessment suggested a possible efficacy signal for invasive treatment relative to conservative treatment but possibly at the risk of increased bleeding events. CONCLUSIONS: There is a paucity of evidence comparing health outcomes between different coronary interventions in frail patients aged 75 years or older with ACS. Available evidence is at high risk of bias. Given the growing importance of ACS in frail patients aged 75 years or older, new studies are needed to inform optimal ACS care for this population. Future studies should rigorously adjust for confounders.

Identifying daily activities of patient work for type 2 diabetes and co-morbidities: a deep learning and wearable camera approach.

OBJECTIVE: People are increasingly encouraged to self-manage their chronic conditions; however, many struggle to practise it effectively. Most studies that investigate patient work (ie, tasks involved in self-management and contexts influencing such tasks) rely on self-reports, which are subject to recall and other biases. Few studies use wearable cameras and deep learning to capture and classify patient work activities automatically. MATERIALS AND METHODS: We propose a deep learning approach to classify activities of patient work collected from wearable cameras, thereby studying self-management routines more effectively. Twenty-six people with type 2 diabetes and comorbidities wore a wearable camera for a day, generating more than 400 h of video across 12 daily activities. To classify these video images, a weighted ensemble network that combines Linear Discriminant Analysis, Deep Convolutional Neural Networks, and Object Detection algorithms is developed. Performance of our model is assessed using Top-1 and Top-5 metrics, compared against manual classification conducted by 2 independent researchers. RESULTS: Across 12 daily activities, our model achieved on average the best Top-1 and Top-5 scores of 81.9 and 86.8, respectively. Our model also outperformed other non-ensemble techniques in terms of Top-1 and Top-5 scores for most activity classes, demonstrating the superiority of leveraging weighted ensemble techniques. CONCLUSIONS: Deep learning can be used to automatically classify daily activities of patient work collected from wearable cameras with high levels of accuracy. Using wearable cameras and a deep learning approach can offer an alternative approach to investigate patient work, one not subjected to biases commonly associated with self-report methods.

Weak label based Bayesian U-Net for optic disc segmentation in fundus images.

Fundus images have been widely used in routine examinations of ophthalmic diseases. For some diseases, the pathological changes mainly occur around the optic disc area; therefore, detection and segmentation of the optic disc are critical pre-processing steps in fundus image analysis. Current machine learning based optic disc segmentation methods typically require manual segmentation of the optic disc for the supervised training. However, it is time consuming to annotate pixel-level optic disc masks and inevitably induces inter-subject variance. To address these limitations, we propose a weak label based Bayesian U-Net exploiting Hough transform based annotations to segment optic discs in fundus images. To achieve this, we build a probabilistic graphical model and explore a Bayesian approach with the state-of-the-art U-Net framework. To optimize the model, the expectation-maximization algorithm is used to estimate the optic disc mask and update the weights of the Bayesian U-Net, alternately. Our evaluation demonstrates strong performance of the proposed method compared to both fully- and weakly-supervised baselines.

Prediction of anxiety disorders using a feature ensemble based bayesian neural network.

Anxiety disorders are common among youth, posing risks to physical and mental health development. Early screening can help identify such disorders and pave the way for preventative treatment. To this end, the Youth Online Diagnostic Assessment (YODA) tool was developed and deployed to predict youth disorders using online screening questionnaires filled by parents. YODA facilitated collection of several novel unique datasets of self-reported anxiety disorder symptoms. Since the data is self-reported and often noisy, feature selection needs to be performed on the raw data to improve accuracy. However, a single set of selected features may not be informative enough. Consequently, in this work we propose and evaluate a novel feature ensemble based Bayesian Neural Network (FE-BNN) that exploits an ensemble of features for improving the accuracy of disorder predictions. We evaluate the performance of FE-BNN on three disorder-specific datasets collected by YODA. Our method achieved the AUC of 0.8683, 0.8769, 0.9091 for the predictions of Separation Anxiety Disorder, Generalized Anxiety Disorder and Social Anxiety Disorder, respectively. These results provide initial evidence that our method outperforms the original diagnostic scoring function of YODA and several other baseline methods for three anxiety disorders, which can practically help prioritizing diagnostic interviews. Our promising results call for investigation of interpretable methods maintaining high predictive accuracy.

Analyzing security issues of android mobile health and medical applications.

OBJECTIVE: We conduct a first large-scale analysis of mobile health (mHealth) apps available on Google Play with the goal of providing a comprehensive view of mHealth apps' security features and gauging the associated risks for mHealth users and their data. MATERIALS AND METHODS: We designed an app collection platform that discovered and downloaded more than 20 000 mHealth apps from the Medical and Health & Fitness categories on Google Play. We performed a suite of app code and traffic measurements to highlight a range of app security flaws: certificate security, sensitive or unnecessary permission requests, malware presence, communication security, and security-related concerns raised in user reviews. RESULTS: Compared to baseline non-mHealth apps, mHealth apps generally adopt more reliable signing mechanisms and request fewer dangerous permissions. However, significant fractions of mHealth apps expose users to serious security risks. Specifically, 1.8% of mHealth apps package suspicious codes (eg, trojans), 45.0% rely on unencrypted communication, and as much as 23.0% of personal data (eg, location information and passwords) is sent on unsecured traffic. An analysis of the app reviews reveals that mHealth app users are largely unaware of the surfaced security issues. CONCLUSION: Despite being better aligned with security best practices than non-mHealth apps, mHealth apps are still far from ensuring robust security guarantees. App users, clinicians, technology developers, and policy makers alike should be cognizant of the uncovered security issues and weigh them carefully against the benefits of mHealth apps.

Single cell morphology distinguishes genotype and drug effect in Hereditary Spastic Paraplegia.

A central need for neurodegenerative diseases is to find curative drugs for the many clinical subtypes, the causative gene for most cases being unknown. This requires the classification of disease cases at the genetic and cellular level, an understanding of disease aetiology in the subtypes and the development of phenotypic assays for high throughput screening of large compound libraries. Herein we describe a method that facilitates these requirements based on cell morphology that is being increasingly used as a readout defining cell state. In patient-derived fibroblasts we quantified 124 morphological features in 100,000 cells from 15 people with two genotypes (SPAST and SPG7) of Hereditary Spastic Paraplegia (HSP) and matched controls. Using machine learning analysis, we distinguished between each genotype and separated them from controls. Cell morphologies changed with treatment with noscapine, a tubulin-binding drug, in a genotype-dependent manner, revealing a novel effect on one of the genotypes (SPG7). These findings demonstrate a method for morphological profiling in fibroblasts, an accessible non-neural cell, to classify and distinguish between clinical subtypes of neurodegenerative diseases, for drug discovery, and potentially for biomarkers of disease severity and progression.

Mobile health and privacy: cross sectional study.

OBJECTIVES: To investigate whether and what user data are collected by health related mobile applications (mHealth apps), to characterise the privacy conduct of all the available mHealth apps on Google Play, and to gauge the associated risks to privacy. DESIGN: Cross sectional study SETTING: Health related apps developed for the Android mobile platform, available in the Google Play store in Australia and belonging to the medical and health and fitness categories. PARTICIPANTS: Users of 20 991 mHealth apps (8074 medical and 12 917 health and fitness found in the Google Play store: in-depth analysis was done on 15 838 apps that did not require a download or subscription fee compared with 8468 baseline non-mHealth apps. MAIN OUTCOME MEASURES: Primary outcomes were characterisation of the data collection operations in the apps code and of the data transmissions in the apps traffic; analysis of the primary recipients for each type of user data; presence of adverts and trackers in the app traffic; audit of the app privacy policy and compliance of the privacy conduct with the policy; and analysis of complaints in negative app reviews. RESULTS: 88.0% (n=18 472) of mHealth apps included code that could potentially collect user data. 3.9% (n=616) of apps transmitted user information in their traffic. Most data collection operations in apps code and data transmissions in apps traffic involved external service providers (third parties). The top 50 third parties were responsible for most of the data collection operations in app code and data transmissions in app traffic (68.0% (2140), collectively). 23.0% (724) of user data transmissions occurred on insecure communication protocols. 28.1% (5903) of apps provided no privacy policies, whereas 47.0% (1479) of user data transmissions complied with the privacy policy. 1.3% (3609) of user reviews raised concerns about privacy. CONCLUSIONS: This analysis found serious problems with privacy and inconsistent privacy practices in mHealth apps. Clinicians should be aware of these and articulate them to patients when determining the benefits and risks of mHealth apps.

Deep convolutional neural networks based ECG beats classification to diagnose cardiovascular conditions.

Medical practitioners need to understand the critical features of ECG beats to diagnose and identify cardiovascular conditions accurately. This would be greatly facilitated by identifying the significant features of frequency components in temporal ECG wave-forms using computational methods. In this study, we have proposed a novel ECG beat classifier based on a customized VGG16-based Convolution Neural Network (CNN) that uses the time-frequency representation of temporal ECG, and a method to identify the contribution of interpretable ECG frequencies when classifying based on the SHapley Additive exPlanations (SHAP) values. We applied our model to the MIT-BIH arrhythmia dataset to classify the ECG beats and to characterise of the beats frequencies. This model was evaluated with two advanced time-frequency analysis methods. Our results indicated that for 2-4 classes our proposed model achieves a classification accuracy of 100% and for 5 classes it achieves a classification accuracy of 99.90%. We have also tested the proposed model using premature ventricular contraction beats from the American Heart Association (AHA) database and normal beats from Lobachevsky University Electrocardiography database (LUDB) and obtained a classification accuracy of 99.91% for the 5-classes case. In addition, SHAP value increased the interpretability of the ECG frequency features. Thus, this model could be applicable to the automation of the cardiovascular diagnosis system and could be used by clinicians.

Benchmarking Audio Signal Representation Techniques for Classification with Convolutional Neural Networks.

Audio signal classification finds various applications in detecting and monitoring health conditions in healthcare. Convolutional neural networks (CNN) have produced state-of-the-art results in image classification and are being increasingly used in other tasks, including signal classification. However, audio signal classification using CNN presents various challenges. In image classification tasks, raw images of equal dimensions can be used as a direct input to CNN. Raw time-domain signals, on the other hand, can be of varying dimensions. In addition, the temporal signal often has to be transformed to frequency-domain to reveal unique spectral characteristics, therefore requiring signal transformation. In this work, we overview and benchmark various audio signal representation techniques for classification using CNN, including approaches that deal with signals of different lengths and combine multiple representations to improve the classification accuracy. Hence, this work surfaces important empirical evidence that may guide future works deploying CNN for audio signal classification purposes.

Brain volumetric and fractal analysis of synthetic MRI: A comparative study with conventional 3D T1-weighted images.

PURPOSE: The estimation of brain volumetric measurements based on Synthetic MRI (SyMRI) is easy and fast, however, the consistency of brain volumetric and morphologic measurements based on SyMRI and 3D T1WI should be further addressed. The current study evaluated the impact of spatial resolution on brain volumetric and morphologic measurements using SyMRI, and test whether the brain measurements derived from SyMRI were consistent with those resulted from 3D T1WI. METHOD: Brain volumetric and fractal analysis were applied to thirty healthy subjects, each underwent four SyMRI acquisitions with different spatial resolutions (1 × 1 × 2 mm, 1 × 1x3mm, 1 × 1 × 4 mm, 2 × 2 × 2 mm) and a 3D T1WI (1 × 1 × 1 mm isotropic). The consistency of the SyMRI measurements was tested using one-way non-parametric Kruskal-Wallis test and post hoc Dwass-Steel-Critchlow-Fligner test. The association between SyMRI and 3D T1WI derived measurements was evaluated using linear regression models. RESULTS: Our results demonstrated that both in- and through-plane resolutions show an impact on brain volumetric measurements, while brain parenchymal volume showed high consistency across the SyMRI acquisitions, and high association with the measurements from 3D T1WI. In addition, SyMRI with 1 × 1 × 4 mm resolution showed the strongest association with 3D T1WI compared to other SyMRI acquisitions in both volumetric and fractal analyses. Moreover, substantial differences were found in fractal dimension of both gray and white matter between the SyMRI and 3D T1WI tissue segmentations. CONCLUSIONS: Our results suggested that the measurements from SyMRI with relatively higher in-plane and lower through-plane resolution (1 × 1 × 4 mm) are much closer to 3D T1WI.