Simulating doctors' thinking logic for chest X-ray report generation via Transformer-based Semantic Query learning.

Medical report generation can be treated as a process of doctors' observing, understanding, and describing images from different perspectives. Following this process, this paper innovatively proposes a Transformer-based Semantic Query learning paradigm (TranSQ). Briefly, this paradigm is to learn an intention embedding set and make a semantic query to the visual features, generate intent-compliant sentence candidates, and form a coherent report. We apply a bipartite matching mechanism during training to realize the dynamic correspondence between the intention embeddings and the sentences to induct medical concepts into the observation intentions. Experimental results on two major radiology reporting datasets (i.e., IU X-ray and MIMIC-CXR) demonstrate that our model outperforms state-of-the-art models regarding generation effectiveness and clinical efficacy. In addition, comprehensive ablation experiments fully validate the TranSQ model's innovation and interpretation. The code is available at https://github.com/zjukongming/TranSQ.

Learning Representations from Medical Text for Effective Diagnoses and Knowledge Discovery.

Discovering knowledge and effectively predicting target events are two main goals of medical text mining. However, few models can achieve them simultaneously. In this study, we investigated the possibility of discovering knowledge and predicting diagnosis at once via raw medical text. We proposed the Enhanced Neural Topic Model (ENTM), a variant of the neural topic model, to learn interpretable representations. We introduced the auxiliary loss set to improve the effectiveness of learned representations. Then, we used learned representations to train a softmax regression model to predict target events. As each element in representations learned by the ENTM has an explicit semantic meaning, weights in softmax regression represent potential knowledge of whether an element is a significant factor in predicting diagnosis. We adopted two independent medical text datasets to evaluate our ENTM model. Results indicate that our model performed better than the latest pretrained neural language models. Meanwhile, analysis of model parameters indicates that our model has the potential discover knowledge from data.Clinical relevance- This work provides a model that can effectively predict patient diagnosis and has the potential to discover knowledge from medical text.

Identifying underlying patterns in Alzheimer's disease trajectory: a deep learning approach and Mendelian randomization analysis.

Background: Alzheimer's disease (AD) is a heterogeneously progressive neurodegeneration disorder with varied rates of deterioration, either between subjects or within different stages of a certain subject. Estimating the course of AD at early stages has treatment implications. We aimed to analyze disease progression to identify distinct patterns in AD trajectory. Methods: We proposed a deep learning model to identify underlying patterns in the trajectory from cognitively normal (CN) to a state of mild cognitive impairment (MCI) to AD dementia, by jointly predicting time-to-conversion and clustering out distinct subgroups characterized by comprehensive features as well as varied progression rates. We designed and validated our model on the ADNI dataset (1370 participants). Prediction of time-to-conversion in AD trajectory was used to validate the expression of the identified patterns. Causality between patterns and time-to-conversion was further inferred using Mendelian randomization (MR) analysis. External validation was performed on the AIBL dataset (233 participants). Findings: The proposed model clustered out patterns characterized by significantly different biomarkers and varied progression rates. The discovered patterns also showed a strong prediction ability, as indicated by hazard ratio (CN→MCI, HR = 3.51, p < 0.001; MCI→AD, HR = 8.11, p < 0.001), C-Index (CN→MCI, 0.618; MCI→AD, 0.718), and AUC (CN→MCI, 3 years 0.802, 5 years 0.876; MCI→AD, 3 years 0.914, 5 years 0.957). In the external validation cohort, our model demonstrated competitive performance on conversion time prediction (CN→MCI, C-Index = 0.693; MCI→AD, C-Index = 0.752). Moreover, suggestive associations between CN→MCI/MCI→AD patterns with four/three SNPs were mediated and MR analysis indicated a causal link between MCI→AD patterns and time-to-conversion in the first three years. Interpretation: Our proposed model identifies biologically and clinically meaningful patterns from real-world data and provides promising performance on time-to-conversion prediction in AD trajectory, which could promote the understanding of disease progression, facilitate clinical trial design, and provide potential for decision-making. Funding: The National Key Research and Development Program of China, the Key R&D Program of Zhejiang, and the National Nature Science Foundation of China.

A knowledge-enhanced transform-based multimodal classifier for microbial keratitis identification.

Microbial keratitis, a nonviral corneal infection caused by bacteria, fungi, and protozoa, is an urgent condition in ophthalmology requiring prompt treatment in order to prevent severe complications of corneal perforation and vision loss. It is difficult to distinguish between bacterial and fungal keratitis from image unimodal alone, as the characteristics of the sample images themselves are very close. Therefore, this study aims to develop a new deep learning model called knowledge-enhanced transform-based multimodal classifier that exploited the potential of slit-lamp images along with treatment texts to identify bacterial keratitis (BK) and fungal keratitis (FK). The model performance was evaluated in terms of the accuracy, specificity, sensitivity and the area under the curve (AUC). 704 images from 352 patients were divided into training, validation and testing set. In the testing set, our model reached the best accuracy was 93%, sensitivity was 0.97(95% CI [0.84,1]), specificity was 0.92(95% CI [0.76,0.98]) and AUC was 0.94(95% CI [0.92,0.96]), exceeding the benchmark accuracy of 0.86. The diagnostic average accuracies of BK ranged from 81 to 92%, respectively and those for FK were 89-97%. It is the first study to focus on the influence of disease changes and medication interventions on infectious keratitis and our model outperformed the benchmark models and reaching the state-of-the-art performance.

Comparative Effectiveness of Digital Cognitive Behavioral Therapy vs Medication Therapy Among Patients With Insomnia.

Importance: Although digital cognitive behavioral therapy for insomnia (dCBT-I) has been studied in many randomized clinical trials and is recommended as a first-line treatment option, few studies have systematically examined its effectiveness, engagement, durability, and adaptability in clinical settings. Objective: To evaluate the clinical effectiveness, engagement, durability, and adaptability of dCBT-I. Design, Setting, and Participants: This retrospective cohort study was conducted using longitudinal data collected via a mobile app named Good Sleep 365 between November 14, 2018, and February 28, 2022. Three therapeutic modes (ie, dCBT-I, medication, and their combination) were compared at month 1, month 3, and month 6 (primary). Inverse probability of treatment weighting (IPTW) using propensity scores was applied to enable homogeneous comparisons between the 3 groups. Exposures: Treatment with dCBT-I, medication therapy, or combination therapy according to prescriptions. Main Outcomes and Measures: The Pittsburgh Sleep Quality Index (PSQI) score and its essential subitems were used as the primary outcomes. Effectiveness on comorbid somnolence, anxiety, depression, and somatic symptoms were used as secondary outcomes. Cohen d effect size, P value, and standardized mean difference (SMD) were used to measure differences in treatment outcomes. Changes in outcomes and response rates (≥3 points change in PSQI score) were also reported. Results: A total of 4052 patients (mean [SD] age, 44.29 [12.01] years; 3028 [74.7%] female participants) were selected for dCBT-I (n = 418), medication (n = 862), and their combination (n = 2772). Compared with the change in PSQI score at 6 months for participants receiving medication alone (from a mean [SD] of 12.85 [3.49] to 8.92 [4.03]), both dCBT-I (from a mean [SD] of 13.51 [3.03] to 7.15 [3.25]; Cohen d, -0.50; 95% CI, -0.62 to -0.38; P < .001; SMD = 0.484) and combination therapy (from a mean [SD] of 12.92 [3.49] to 6.98 [3.43]; Cohen d, 0.50; 95% CI, 0.42 to 0.58; P < .001; SMD = 0.518) were associated with significant reductions; dCBT-I had a comparable effect as combination therapy (Cohen d, 0.05; 95% CI, -0.05 to 0.15; P = .66; SMD = 0.05), but showed unstable durability. Outcomes of dCBT-I improved steadily and rapidly during the first 3 months, and then fluctuated. The response rates with dCBT-I and combination therapy were higher than with medication. Changes in secondary outcomes indicated statistically significant benefits from dCBT-I and combination therapy. The results of subgroup analysis were consistent with the main findings, demonstrating the superiority of dCBT-I vs medication therapy in various subpopulations. Conclusions and Relevance: In this study, clinical evidence suggested that combination therapy was optimal, and dCBT-I was more effective than medication therapy, with long-term benefits for insomnia. Future studies are needed to analyze its clinical effectiveness and reliability in distinct subpopulations.

Knowledge-aware patient representation learning for multiple disease subtypes.

Learning latent representations of patients with a target disease is a core problem in a broad range of downstream applications, such as clinical endpoint prediction. The suffering of patients may have multiple subtypes with certain similarities and differences, which need to be addressed for learning effective patient representation to facilitate the downstream tasks. However, existing studies either ignore the distinction of disease subtypes to learn disease-level representations, or neglect the correlations between subtypes and only learn disease subtype-level representations, which affects the performance of patient representation learning. To alleviate this problem, we studied how to effectively integrate data from all disease subtypes to improve the representation of each subtype. Specifically, we proposed a knowledge-aware shared-private neural network model to explicitly use disease-oriented knowledge and learn shared and specific representations from the disease and its subtype perspectives. To evaluate the feasibility of the proposed model, we conducted a particular downstream task, i.e., clinical endpoint prediction, on the basis of the learned patient presentations. The results on the real-world clinical datasets demonstrated that our model could yield a significant improvement over state-of-the-art models.

Adversarial reinforcement learning for dynamic treatment regimes.

Treatment recommendation, as a critical task of delivering effective interventions according to patient state and expected outcome, plays a vital role in precision medicine and healthcare management. As a well-suited tactic to learn optimal policies of recommender systems, reinforcement learning is promising to address the challenge of treatment recommendation. However, existing solutions mostly require frequent interactions between treatment recommender systems and clinical environment, which are expensive, time-consuming, and even infeasible in clinical practice. In this study, we present a novel model-based offline reinforcement learning approach to optimize a treatment policy by utilizing patient treatment trajectories in Electronic Health Records (EHRs). Specifically, a patient treatment trajectory simulator is firstly constructed based on the ground-truth trajectories in EHRs. Thereafter, the constructed simulator is utilized to model the online interactions between the treatment recommender system and clinical environment. In this way, the counterfactual trajectories can be generated. To alleviate the bias deriving from the ground-truth and the counterfactual trajectories, an adversarial network is incorporated into the proposed model, such that a large space of treatment actions can be explored with the scaled rewards. The proposed model is evaluated on a simulated dataset and a real-world dataset. The experimental results demonstrate that the proposed model is superior to other methods, giving rise to a new solution for dynamic treatment regimes and beyond.

Identifying Patients With Heart Failure Who Are Susceptible to De Novo Acute Kidney Injury: Machine Learning Approach.

BACKGROUND: Studies have shown that more than half of patients with heart failure (HF) with acute kidney injury (AKI) have newonset AKI, and renal function evaluation markers such as estimated glomerular filtration rate are usually not repeatedly tested during the hospitalization. As an independent risk factor, delayed AKI recognition has been shown to be associated with the adverse events of patients with HF, such as chronic kidney disease and death. OBJECTIVE: The aim of this study is to develop and assess of an unsupervised machine learning model that identifies patients with HF and normal renal function but who are susceptible to de novo AKI. METHODS: We analyzed an electronic health record data set that included 5075 patients admitted for HF with normal renal function, from which 2 phenogroups were categorized using an unsupervised machine learning algorithm called K-means clustering. We then determined whether the inferred phenogroup index had the potential to be an essential risk indicator by conducting survival analysis, AKI prediction, and the hazard ratio test. RESULTS: The AKI incidence rate in the generated phenogroup 2 was significantly higher than that in phenogroup 1 (group 1: 106/2823, 3.75%; group 2: 259/2252, 11.50%; P<.001). The survival rate of phenogroup 2 was consistently lower than that of phenogroup 1 (P<.005). According to logistic regression, the univariate model using the phenogroup index achieved promising performance in AKI prediction (sensitivity 0.710). The generated phenogroup index was also significant in serving as a risk indicator for AKI (hazard ratio 3.20, 95% CI 2.55-4.01). Consistent results were yielded by applying the proposed model on an external validation data set extracted from Medical Information Mart for Intensive Care (MIMIC) III pertaining to 1006 patients with HF and normal renal function. CONCLUSIONS: According to a machine learning analysis on electronic health record data, patients with HF who had normal renal function were clustered into separate phenogroups associated with different risk levels of de novo AKI. Our investigation suggests that using machine learning can facilitate patient phengrouping and stratification in clinical settings where the identification of high-risk patients has been challenging.

Disentangled representation for sequential treatment effect estimation.

BACKGROUND AND OBJECTIVE: Treatment effect estimation, as a fundamental problem in causal inference, focuses on estimating the outcome difference between different treatments. However, in clinical observational data, some patient covariates (such as gender, age) not only affect the outcomes but also affect the treatment assignment. Such covariates, named as confounders, produce distribution discrepancies between different treatment groups, thereby introducing the selection bias for the estimation of treatment effects. The situation is even more complicated in longitudinal data, because the confounders are time-varying that are subject to patient history and meanwhile affect the future outcomes and treatment assignments. Existing methods mainly work on cross-sectional data obtained at a specific time point, but cannot process the time-varying confounders hidden in the longitudinal data. METHODS: In this study, we address this problem for the first time by disentangled representation learning, which considers the observational data as consisting of three components, including outcome-specific factors, treatment-specific factors, and time-varying confounders. Based on this, the proposed approach adopts a recurrent neural network-based framework to process sequential information and learn the disentangled representations of the components from longitudinal observational sequences, captures the posterior distributions of latent factors by multi-task learning strategy. Moreover, mutual information-based regularization is adopted to eliminate the time-varying confounders. In this way, the association between patient history and treatment assignment is removed and the estimation can be effectively conducted. RESULTS: We evaluate our model in a realistic set-up using a model of tumor growth. The proposed model achieves the best performance over benchmark models for both one-step ahead prediction (0.70% vs 0.74% for the-state-of-the-art model, when γ = 3. Measured by normalized root mean square error, the lower the better) and five-step ahead prediction (1.47% vs 1.83%) in most cases. By increasing the effect of confounders, our proposed model always shows superiority against the state-of-the-art model. In addition, we adopted T-SNE to visualize the disentangled representations and present the effectiveness of disentanglement explicitly and intuitively. CONCLUSIONS: The experimental results indicate the powerful capacity of our model in learning disentangled representations from longitudinal observational data and dealing with the time-varying confounders, and demonstrate the surpassing performance achieved by our proposed model on dynamic treatment effect estimation.

Attribute-aware interpretation learning for thyroid ultrasound diagnosis.

Thyroid nodule diagnosis from ultrasound images is a critical computer-aided diagnosis task. Previous works tried to imitate the doctor's diagnosis logic by considering the key attributes to improve the diagnosis performance and explaining the conclusion. However, their clinical feasibilities are still ambiguous because of the ignorance of the correlation between attribute features and global characteristics, as well as the lack of clinical effectiveness evaluation of result interpretations. Following the common logic of ultrasonic investigation, we design a novel Attribute-Aware Interpretation Learning (AAIL) model, consisting of attribute properties discovery module and attribute-global feature fusion module. Adequate result interpretation ensures reliability and transparency of diagnostic conclusions, including the visualization of attribute features and the relationship between attributes and the global feature. Extensive experiments on a practical dataset demonstrate the model's effectiveness, and an innovative human-computer collaborative experiment demonstrates the auxiliary diagnostic ability of the interpretations that can benefit professional doctors.

CD-Surv: a contrastive-based model for dynamic survival analysis.

Survival analysis, aimed at investigating the relationships between covariates and event time, has exhibited profound effects on health service management. Longitudinal data with sequential patterns, such as electronic health records (EHRs), contain a large volume of patient treatment trajectories, and therefore, provide great potential for survival analysis. However, most existing studies address the survival analysis problem in a static manner, that is, they only utilize a fraction of longitudinal data, ignore the correlations between multiple visits, and usually may not be able to capture the latent representations of patient treatment trajectories. This inevitably deteriorates the performance of the survival analysis. To address this challenge, we propose an end-to-end contrastive-based model CD-Surv to better understand the patient treatment trajectories and dynamically predict the survival probability of a target patient. Specifically, two data augmentation strategies, namely, mask generation and shuffle generation, are adopted to augment the real treatment trajectories documented in the EHR. Based on this, the hidden representations of the real trajectories can be improved by utilizing contrastive learning between augmented and real trajectories. We evaluated our proposed CD-Surv on two real-world datasets, and the experimental results indicated that our proposed model could outperform state-of-the-art baselines on various evaluation metrics.

Comparing Machine Learning Models and Statistical Models for Predicting Heart Failure Events: A Systematic Review and Meta-Analysis.

Objective: To compare the performance, clinical feasibility, and reliability of statistical and machine learning (ML) models in predicting heart failure (HF) events. Background: Although ML models have been proposed to revolutionize medicine, their promise in predicting HF events has not been investigated in detail. Methods: A systematic search was performed on Medline, Web of Science, and IEEE Xplore for studies published between January 1, 2011 to July 14, 2021 that developed or validated at least one statistical or ML model that could predict all-cause mortality or all-cause readmission of HF patients. Prediction Model Risk of Bias Assessment Tool was used to assess the risk of bias, and random effect model was used to evaluate the pooled c-statistics of included models. Result: Two-hundred and two statistical model studies and 78 ML model studies were included from the retrieved papers. The pooled c-index of statistical models in predicting all-cause mortality, ML models in predicting all-cause mortality, statistical models in predicting all-cause readmission, ML models in predicting all-cause readmission were 0.733 (95% confidence interval 0.724-0.742), 0.777 (0.752-0.803), 0.678 (0.651-0.706), and 0.660 (0.633-0.686), respectively, indicating that ML models did not show consistent superiority compared to statistical models. The head-to-head comparison revealed similar results. Meanwhile, the immoderate use of predictors limited the feasibility of ML models. The risk of bias analysis indicated that ML models' technical pitfalls were more serious than statistical models'. Furthermore, the efficacy of ML models among different HF subgroups is still unclear. Conclusions: ML models did not achieve a significant advantage in predicting events, and their clinical feasibility and reliability were worse.

Deep-CSA: Deep Contrastive Learning for Dynamic Survival Analysis With Competing Risks.

Survival analysis (SA) is widely used to analyze data in which the time until the event is of interest. Conventional SA techniques assume a specific form for viewing the distribution of survival time as the hitting time of a stochastic process, and explicitly model the relationship between covariates and the distribution of the events hitting time. Although valuable, existing SA models seldom consider to model the dynamic correlations between covariates and more than one event of interest (i.e., competing risks) in the disease progression of subjects. To alleviate this critical problem, we propose a novel deep contrastive learning model to obtain a deep understanding of disease progression of subjects with competing risks from their longitudinal observational data. Specifically, we design a self-supervised objective for learning dynamic representations of subjects suffering from multiple competing risks, such that the relationship between covariates and each specific competing risk changes over time can be well captured. Experiments on two open-source clinical datasets, i.e., MIMIC-III and EICU, demonstrate the effectiveness of our proposed model, with remarkable improvements over the state-of-the-art SA models.

A Unified Machine Reading Comprehension Framework for Cohort Selection.

Cohort selection is an essential prerequisite for clinical research, determining whether an individual satisfies given selection criteria. Previous works for cohort selection usually treated each selection criterion independently and ignored not only the meaning of each selection criterion but the relations among cohort selection criteria. To solve the problems above, we propose a novel unified machine reading comprehension (MRC) framework. In this MRC framework, we design simple rules to generate questions for each criterion from cohort selection guidelines and treat clues extracted by trigger words from patients' medical records as passages. A series of state-of-the-art MRC models based on BiDAF, BIMPM, BERT, BioBERT, NCBI-BERT, and RoBERTa are deployed to determine which question and passage pairs match. We also introduce a cross-criterion attention mechanism on representations of question and passage pairs to model relations among cohort selection criteria. Results on two datasets, that is, the dataset of the 2018 National NLP Clinical Challenge (N2C2) for cohort selection and a dataset from the MIMIC-III dataset, show that our NCBI-BERT MRC model with cross-criterion attention mechanism achieves the highest micro-averaged F1-score of 0.9070 on the N2C2 dataset and 0.8353 on the MIMIC-III dataset. It is competitive to the best system that relies on a large number of rules defined by medical experts on the N2C2 dataset. Comparing these two models, we find that the NCBI-BERT MRC model mainly performs worse on mathematical logic criteria. When using rules instead of the NCBI-BERT MRC model on some criteria regarding mathematical logic on the N2C2 dataset, we obtain a new benchmark with an F1-score of 0.9163, indicating that it is easy to integrate rules into MRC models for improvement.

On learning disentangled representations for individual treatment effect estimation.

OBJECTIVE: Estimating the individualized treatment effect (ITE) from observational data is a challenging task due to selection bias, which results from the distributional discrepancy between different treatment groups caused by the dependence between features and assigned treatments. This dependence is induced by the factors related to the treatment assignment. We hypothesize that features consist of three types of latent factors: outcome-specific factors, treatment-specific factors and confounders. Then, we aim to reduce the influence of treatment-related factors, i.e., treatment-specific factors and confounders, on outcome prediction to mitigate the effects of selection bias. METHOD: We present a novel representation learning model in which both the main task of outcome prediction and the auxiliary task of classifying the treatment assignment are used to learn the outcome-oriented and treatment-oriented latent representations, respectively. However, since the confounders are related to both treatment assignment and outcome, it is still contained in the representations. To further reduce influence of the confounders contained in both representations, individualized orthogonal regularization is incorporated into the proposed model. The orthogonal regularization forces the outcome-oriented and treatment-oriented latent representations of an individual to be vertical in the inner product space, meaning they are orthogonal with each other, and the common information of confounder is reduced. Such that the ITE can be estimated more precisely without the effects of selection bias. RESULT: We evaluate our proposed model on a semi-simulated dataset and a real-world dataset. The experimental results demonstrate that the proposed model achieves competitive or better performance compared with the performances of the state-of-the-art models. CONCLUSION: The proposed method is well performed on ITE estimation with the ability to reduce selection bias thoroughly by incorporating an auxiliary task and adopting orthogonal regularization to disentangle the latent factors. SIGNIFICANCE: This paper offers a novel method of reducing selection bias in estimating the ITE from observational data by disentangled representation learning.

Health Natural Language Processing: Methodology Development and Applications.

With the rapid growth of information technology, the necessity for processing substantial amounts of health data using advanced information technologies is increasing. A large amount of valuable data exists in natural text such as diagnosis text, discharge summaries, online health discussions, and eligibility criteria of clinical trials. Health natural language processing, as an interdisciplinary field of natural language processing and health care, plays a substantial role in a wide scope of both methodology development and applications. This editorial shares the most recent methodology innovations of health natural language processing and applications in the medical domain published in this JMIR Medical Informatics special theme issue entitled "Health Natural Language Processing: Methodology Development and Applications".

Towards Predictive Analysis on Disease Progression: A Variational Hawkes Process Model.

Massively available longitudinal data about long-term disease trajectories of patients provides a golden mine for the understanding of disease progression and efficient health service delivery. It calls for quantitative modeling of disease progression, which is a tricky problem due to the complexity of the disease progression process as well as the irregularity of time documented in trajectories. In this study, we tackle the problem with the goal of predictively analyzing disease progression. Specifically, we propose a novel Variational Hawkes Process (VHP) model to generalize disease progression and predict future patient states based on the clinical observational data of past disease trajectories. First, Hawkes Process captures the intensity of irregular visits in a trajectory documented to medical facilities and controls the aforementioned information flowing into future visits. Thereafter, the captured intensity is incorporated into a Variational Auto-Encoder to generate the representation of the future partial disease trajectory for a target patient in a predictive manner. To further improve the prediction performance, we equip the proposed model with a disease trajectory discriminator to distinguish the generated trajectories from real ones. We evaluate the proposed model on two public datasets from the MIMIC-III database pertaining to heart failure and sepsis patients, respectively, and one real-world dataset from a Chinese hospital pertaining to heart failure patients with multiple admissions. Experimental results demonstrate that the proposed model significantly outperforms state-of-the-art baselines, and may derive a set of practical implications that can benefit a wide spectrum of management and applications on disease progression.