Clinical decision support system (CDSS) plays an essential role nowadays and CDSS for treatment provides clinicians with the clinical evidence of candidate prescriptions to assist them in making patient-specific decisions. Therefore, it is essential to find a partition of patients such that patients with similar clinical conditions are grouped together and the preferred prescriptions for different groups are diverged. A comprehensive clinical guideline often provides information of patient partition. However, for most diseases, the guideline is not so detailed that only limited circumstances are covered. This makes it challenging to group patients properly. Here we proposed an approach that combines clinical guidelines with medical data to construct a nested decision tree for patient partitioning and treatment recommendation. Compared with pure data-driven decision tree, the recommendations generated by our model have better guideline adherence and interpretability. The approach was successfully applied in a real-world case study of patients with hyperthyroidism.

Disease diagnosis plays an important role in the application of clinical decision system. When applying artificial intelligence models in disease diagnosis, one should be aware of the spurious correlations learned from the data. Such correlations could be introduced by the biased distribution of the data and would be reinforced during the optimization process. The spurious correlations will cause the model ultra-sensitive to some trivial variations in the input, and even introduce some significant errors when applied in non-independent and identically distributed (non-i.i.d) data with the training data. In this paper, we addressed this problem by applying the global balancing method to adjust the distribution balance in the learned representation space of the text via the assignment of global sample weights. By learning the global sample weights that minimized the balance loss, the text representation space had been iteratively optimized by attenuating the strength of the spurious correlations introduced by the training data. Our algorithm, TCGBR (Text Classification with the Global Balance Regularizer), showed the reduced sensitivity to the trivial variations of the input text and increased diagnosis accuracy when tested on the non-i.i.d data. Furthermore, we built the causal graph to investigate the causal relationships in established diagnostic models and found the diagnostic logic of TCGBR was more in line with the clinical common sense. The presented architecture can be used by practitioners to improve the stability of diagnostic models and the technique of global balancing can be naturally generalized to other representation-based deep learning models.

Accurate diagnosis plays an important role in the clinical decision-making process. To reduce diagnostic errors, the Diagnostic Decision Support System (DDSS) is designed to assist clinicians with the diagnostic decision-making. Recently, DDSS is developed by diagnosis classification models which employ deep learning networks training on real-world patient-visit data which are often in the form of medical codes. We realize the importance of medical code descriptions for diagnosis classification and believe that incorporating medical code descriptions can improve diagnostic accuracy. Incorporating medical code descriptions by mapping medical codes to textual descriptions, we build diagnosis classification models using 3 text classification methods (TextCNN, TextRCNN and CharCNN) on Ambulatory Health Care (AHC) data with 629,827 patient-visits in the United States, to provide diagnostic decision support under the certain situation when patients' data are recorded in the form of medical codes. The experiments demonstrate the improved accuracy performance and time efficiency of our models compared the baseline models (XGBoost and MLP) taking only medical codes as the input.

Objective: Although guidelines have recommended standardized drug treatment for heart failure (HF), there are still many challenges in making the correct clinical decisions due to the complicated clinical situations of HF patients. Each patient would satisfy several recommendations, meaning the decision tree of HF treatment should be nonmutually exclusive, and the same patient would be allocated to several leaf nodes in the decision tree. In the current study, we aim to propose a way to ensemble a nonmutually exclusive decision tree for recommendation system for complicated diseases, such as HF. Methods: The nonmutually exclusive decision tree was constructed via knowledge rules summarized from the HF clinical guidelines. Then similar patients were defined as those who followed the same pattern of leaf node allocation according to the decision tree. The frequent medication patterns for each similar patient were mined using the Apriori algorithms, and we also carried out the outcome prognosis analyses to show the capability for the evidence-based medication recommendations of our nonmutually exclusive decision tree. Results: Based on a large database that included 29,689 patients with 84,705 admissions, we tested the framework for HF treatment recommendation. In the constructed decision tree, the HF treatment recommendations were grouped into two independent parts. The first part was recommendations for new cases, and the second part was recommendations when patients had different historical medication. There are 14 leaf nodes in our decision tree, and most of the leaf nodes had a guideline adherence of around 90%. We reported the top 10 popular similar patients, which accounted for 32.84% of the whole population. In addition, the multiple outcome prognosis analyses were carried out to assess the medications for one of the subgroups of similar patients. Our results showed even for the subgroup of the same similar patients that no one medication pattern would benefit all outcomes. Conclusion: In the present study, the methodology to construct a nonmutually exclusive decision tree for medication recommendations for HF and its application in CDSS was proposed. Our framework is universal for most diseases and could be generally applied in developing the CDSS for treatment.

Background: Type 2 diabetes mellitus (T2DM) and its related complications represent a growing economic burden for many countries and health systems. Diabetes complications can be prevented through better disease control, but there is a large gap between the recommended treatment and the treatment that patients actually receive. The treatment of T2DM can be challenging because of different comprehensive therapeutic targets and individual variability of the patients, leading to the need for precise, personalized treatment. Objective: The aim of this study was to develop treatment recommendation models for T2DM based on deep reinforcement learning. A retrospective analysis was then performed to evaluate the reliability and effectiveness of the models. Methods: The data used in our study were collected from the Singapore Health Services Diabetes Registry, encompassing 189,520 patients with T2DM, including 6,407,958 outpatient visits from 2013 to 2018. The treatment recommendation model was built based on 80% of the dataset and its effectiveness was evaluated with the remaining 20% of data. Three treatment recommendation models were developed for antiglycemic, antihypertensive, and lipid-lowering treatments by combining a knowledge-driven model and a data-driven model. The knowledge-driven model, based on clinical guidelines and expert experiences, was first applied to select the candidate medications. The data-driven model, based on deep reinforcement learning, was used to rank the candidates according to the expected clinical outcomes. To evaluate the models, short-term outcomes were compared between the model-concordant treatments and the model-nonconcordant treatments with confounder adjustment by stratification, propensity score weighting, and multivariate regression. For long-term outcomes, model-concordant rates were included as independent variables to evaluate if the combined antiglycemic, antihypertensive, and lipid-lowering treatments had a positive impact on reduction of long-term complication occurrence or death at the patient level via multivariate logistic regression. Results: The test data consisted of 36,993 patients for evaluating the effectiveness of the three treatment recommendation models. In 43.3% of patient visits, the antiglycemic medications recommended by the model were concordant with the actual prescriptions of the physicians. The concordant rates for antihypertensive medications and lipid-lowering medications were 51.3% and 58.9%, respectively. The evaluation results also showed that model-concordant treatments were associated with better glycemic control (odds ratio [OR] 1.73, 95% CI 1.69-1.76), blood pressure control (OR 1.26, 95% CI, 1.23-1.29), and blood lipids control (OR 1.28, 95% CI 1.22-1.35). We also found that patients with more model-concordant treatments were associated with a lower risk of diabetes complications (including 3 macrovascular and 2 microvascular complications) and death, suggesting that the models have the potential of achieving better outcomes in the long term. Conclusions: Comprehensive management by combining knowledge-driven and data-driven models has good potential to help physicians improve the clinical outcomes of patients with T2DM; achieving good control on blood glucose, blood pressure, and blood lipids; and reducing the risk of diabetes complications in the long term.

In this paper, we developed a personalized anticoagulant treatment recommendation model for atrial fibrillation (AF) patients based on reinforcement learning (RL) and evaluated the effectiveness of the model in terms of short-term and long-term outcomes. The data used in our work were baseline and follow-up data of 8,540 AF patients with high risk of stroke, enrolled in the Chinese Atrial Fibrillation Registry (CAFR) study during 2011 to 2018. We found that in 64.98% of patient visits, the anticoagulant treatment recommended by the RL model were concordant with the actual prescriptions of the clinicians. Model-concordant treatments were associated with less ischemic stroke and systemic embolism (SSE) event compared with non-concordant ones, but no significant difference on the occurrence rate of major bleeding. We also found that higher proportion of model-concordant treatments were associated with lower risk of death. Our approach identified several high-confidence rules, which were interpreted by clinical experts.

Background: Internet hospitals in China are in great demand due to limited and unevenly distributed health care resources, lack of family doctors, increased burdens of chronic diseases, and rapid growth of the aged population. The COVID-19 epidemic catalyzed the expansion of online health care services. In recent years, internet hospitals have been rapidly developed. Ping An Good Doctor is the largest, national online medical entry point in China and is a widely used platform providing online health care services. Objective: This study aims to give a comprehensive description of the characteristics of the online consultations and inquisitions in Ping An Good Doctor. The analyses tried to answer the following questions: (1) What are the characteristics of the consultations in Ping An Good Doctor in terms of department and disease profiles? (2) Who uses the online health services most frequently? and (3) How is the user experience of the online consultations of Ping An Good Doctor? Methods: A total of 35.3 million consultations and inquisitions over the course of 1 year were analyzed with respect to the distributions of departments and diseases, user profiles, and consulting behaviors. Results: The geographical distribution of the usage of Ping An Good Doctor showed that Shandong (18.4%), Yunnan (15.6%), Shaanxi (7.2%), and Guangdong (5.5%) were the provinces that used it the most; they accounted for 46.6% of the total consultations and inquisitions. In terms of department distribution, we found that gynecology and obstetrics (19.2%), dermatology (17.0%), and pediatrics (14.4%) were the top three departments in Ping An Good Doctor. The disease distribution analysis showed that, except for nondisease-specific consultations, acute upper respiratory infection (AURI) (4.1%), pregnancy (2.8%), and dermatitis (2.4%) were the most frequently consulted diseases. In terms of user profiles, females (60.4%) from 19 to 35 years of age were most likely to seek consultations online, in general. The user behavior analyses showed that the peak times of day for online consultations occurred at 10 AM, 3 PM, and 9 PM. Regarding user experience, 93.0% of users gave full marks following their consultations. For some disease-related health problems, such as AURI, dermatitis, and eczema, the feedback scores were above average. Conclusions: The prevalence of internet hospitals, such as Ping An Good Doctor, illustrated the great demand for online health care services that can go beyond geographical limitations. Our analyses showed that nondisease-specific issues and moderate health problems were much more frequently consulted about than severe clinical conditions. This indicated that internet hospitals played the role of the family doctor, which helped to relieve the stress placed on offline hospitals and facilitated people's lives. In addition, good user experiences, especially regarding disease-related inquisitions, suggested that online health services can help solve health problems. With support from the government and acceptance by the public, online health care services could develop at a fast pace and greatly benefit people's daily lives.

In this paper, we study the problem of leveraging dialogue agents learned from reinforcement learning (RL) that can interact with patients for automatic disease screening. This application requires efficient and effective inquiry of appropriate symptoms to make accurate diagnosis recommendations. Existing studies have tried to use RL to perform both symptom inquiry and diagnosis simultaneously, which needs to deal with a large, heterogeneous action space that affects the learning efficiency and effectiveness. To address the challenge, we propose to leverage the models learned from the dialogue data to customize the settings of the reinforcement learning for more efficient action space exploration. In particular, a supervised diagnosis model is built and involved in the definition of state and reward. We also develop the clustering method to form a hierarchy in the action space. These customizations can make the learning task focus on checking the most relevant symptoms, which effectively boost the confidence of diagnosis. Besides, a novel hierarchical reinforcement learning framework with the pretraining strategy is used to reduce the dimension of action space and help the model to converge. For empirical evaluations, we conduct extensive experiments on both synthetic and real-world datasets. The results have demonstrated the superiority of our approach in diagnostic accuracy and interaction efficiency compared with other baseline methods.

Accurate diagnosis plays an important role in the clinical decision-making process. To reduce diagnostic errors, many researchers develop models to assist physicians in predicting the most probable diagnoses during a patient’s visit. In this paper, we focus on the development of the models for clinical diagnostic decision support using a real-world Electronic Health Record (EHR) dataset with 592,715 patient-visits. We propose a novel diagnosis prediction framework, which is based on Bidirectional Encoder Representations from Transformers (BERT), for the disease classification based on textual clinical notes and age information from EHR data, which differs from the previously proposed models by the way of the construction of the input representation by four special embeddings and the composition of the classification layer. We conduct the experiments on the task of multi-class classification for 1,987 diagnosis codes. The experimental results demonstrate the improved performance of our models compared to the baselines trained by the advanced text classification methods. We also investigate the influence of different types of clinical text and age information on the model performance.

Clinical decision support system (CDSS) plays a significant role nowadays and it assists physicians in making decisions for treatment. Generally based on clinical guideline, the principles of the recommendation are provided and may suggest several candidate medications for similar patient group with certain clinical conditions. However, it is challenging to prioritize these candidates and even refine the guideline to a finer level for patient-specific recommendation. Here we propose a method and system to integrate the clinical knowledge and real-world evidence (RWE) for type 2 diabetes treatment, to enable both standardized and personalized medication recommendation. The RWE is generated by medication effectiveness analysis and subgroup analysis. The knowledge model has been verified by clinical experts from the advanced hospitals. The data verification results show that the medications that are consistent with the method recommendation can lead to better clinical outcome in terms of glycemic control, compared to those inconsistent.

Type 2 diabetes mellitus (T2DM) is a chronic disease that requires continuous treatments. T2DM treatments aim to achieve not only short-term but, more importantly, long-term control of the patient’s glucose to a normal level. We believe Reinforcement Learning (RL) can be an effective approach to learn and further recommend the ideal sequence of treatments that optimize the patient’s long-term outcome. In this paper, we implement RL with the deep Q network. Our RL model learns from a T2DM patient registry dataset that consists of the clinical data of newly diagnosed T2DM patients over a twelve-month period including four follow-ups since their first visit. The RL model is trained to recommend the number of oral antidiabetic drugs and the number of insulins. Our experiments have shown that our RL model can improve the long-term hemoglobin A1c goal achieving rate by up to 15%. This confirms that the model learns good medication patterns that favor the long-term glucose control.

A diagnostic model of general diseases could help general practitioners to decrease misdiagnoses and reduce workload. In this paper, we developed a neural network model that can classify potential diagnoses among 100 selected common diseases based on ambulatory health care data. We propose a novel approach to integrate domain knowledge into neural network training. The evaluation results show our model outperforming the baseline model in terms of knowledge consistency and model generalization.

In clinical practice, many patients may have unknown or missing values for some predictors, causing that the developed risk models cannot be directly applied on these patients. In this paper, we propose an incremental learning approach to apply a developed risk model on new patients with unknown predictor values, which imputes a patient's unknown values based on his/her k-nearest neighbors (k-NN) from the incremental population. We perform a real world case study by developing a risk prediction model of stroke for patients with Type 2 diabetes mellitus from EHR data, and incrementally applying the risk model on a sequence of new patients. The experimental results show that our risk prediction model of stroke has good prediction performance. And the k-nearest neighbors based incremental learning approach for data imputation can gradually increase the prediction performance when the model is applied on new patients.

Clinical decision support systems are information technology systems that assist clinical decision-making tasks, which have been shown to enhance clinical performance. Cluster analysis, which groups similar patients together, aims to separate patient cases into phenotypically heterogenous groups and defining therapeutically homogeneous patient subclasses. Useful as it is, the application of cluster analysis in clinical decision support systems is less reported. Here, we describe the usage of cluster analysis in clinical decision support systems, by first dividing patient cases into similar groups and then providing diagnosis or treatment suggestions based on the group profiles. This integration provides data for clinical decisions and compiles a wide range of clinical practices to inform the performance of individual clinicians. We also include an example usage of the system under the scenario of blood lipid management in type 2 diabetes. These efforts represent a step toward promoting patient-centered care and enabling precision medicine.

Along with the growth of numbers of patients with chronic diseases, personal health self-management becomes critical. The heterogeneity of self-management requirements makes the detail design and implementation of self-management program a non-trivial work. In this paper we address the problem with the Personal Health Advisor (PHA) application by introducing a personal health data flow mechanism, as well as modules including personal health risk assessment, similar patients profiling, and health question answering.

Atrial fibrillation (AF) is a common cardiac rhythm disorder, which increases the risk of ischemic stroke and other thromboembolism (TE). Accurate prediction of TE is highly valuable for early intervention to AF patients. However, the prediction performance of previous TE risk models for AF is not satisfactory. In this study, we used integrated machine learning and data mining approaches to build 2-year TE prediction models for AF from Chinese Atrial Fibrillation Registry data. We first performed data cleansing and imputation on the raw data to generate available dataset. Then a series of feature construction and selection methods were used to identify predictive risk factors, based on which supervised learning methods were applied to build the prediction models. The experimental results show that our approach can achieve higher prediction performance (AUC: 0.71~0.74) than previous TE prediction models for AF (AUC: 0.66~0.69), and identify new potential risk factors as well.

With the rapid growth of clinical data and knowledge, feature construction for clinical analysis becomes increasingly important and challenging. Given a clinical dataset with up to hundreds or thousands of columns, the traditional manual feature construction process is usually too labour intensive to generate a full spectrum of features with potential values. As a result, advanced large-scale data analysis technologies, such as feature selection for predictive modelling, cannot be fully utilized for clinical data analysis. In this paper, we propose an automatic feature construction framework for clinical data analysis, namely, Feature++. It leverages available public knowledge to understand the semantics of the clinical data, and is able to integrate external data sources to automatically construct new features based on predefined rules and clinical knowledge. We demonstrate the effectiveness of Feature++ in a typical predictive modelling use case with a public clinical dataset, and the results suggest that the proposed approach is able to fulfil typical feature construction tasks with minimal dataset specific configurations, so that more accurate models can be obtained from various clinical datasets in a more efficient way.

Recent advances in cloud computing and machine learning made it more convenient for researchers to gain insights from massive healthcare data, while performing analyses on healthcare data in current practice still lacks efficiency for researchers. What's more, collaborating among different researchers and sharing analysis results are challenging issues. In this paper, we developed a practice to make analytics process collaborative and analysis results reproducible by exploiting and extending Jupyter Notebook. After applying this practice in our use cases, we can perform analyses and deliver results with less efforts in shorter time comparing to our previous practice.

Disease risk prediction is highly important for early intervention and treatment, and identification of predictive risk factors is the key point to achieve accurate prediction. In addition to original independent features in a dataset, some interacted features, such as comorbidities and combination therapies, may have non-additive influence on the disease outcome and can also be used in risk prediction to improve the prediction performance. However, it is usually difficult to manually identify the possible interacted risk factors due to the combination explosion of features. In this paper, we propose an automatic approach to identify predictive risk factors with interactions using frequent item set mining and feature selection methods. The proposed approach was applied in the real world case study of predicting ischemic stroke and thromboembolism for atrial fibrillation patients on the Chinese atrial fibrillation registry dataset, and the results show that our approach can not only improve the prediction performance, but also identify the comorbidities and combination therapies that have potential influences on TE occurrence for AF.