Enriching representation learning using 53 million patient notes through human phenotype ontology embedding

The Human Phenotype Ontology (HPO) is a dictionary of >15,000 clinical phenotypic terms with defined semantic relationships, developed to standardize phenotypic analysis. Over the last decade, the HPO has been used to accelerate the implementation of precision medicine into clinical practice. In addition, recent research in representation learning, specifically in graph embedding, has led to notable progress in automated prediction via learned features. Here, we present a novel approach to phenotype representation by incorporating phenotypic frequencies based on 53 million full-text health care notes from >1.5 million individuals. We demonstrate the efficacy of our proposed phenotype embedding technique by comparing our work to existing phenotypic similarity-measuring methods. Using phenotype frequencies in our embedding technique, we are able to identify phenotypic similarities that surpass current computational models. Furthermore, our embedding technique exhibits a high degree of agreement with domain experts' judgment. By transforming complex and multidimensional phenotypes from the HPO format into vectors, our proposed method enables efficient representation of these phenotypes for downstream tasks that require deep phenotyping. This is demonstrated in a patient similarity analysis and can further be applied to disease trajectory and risk prediction.

A unified graph model based on molecular data binning for disease subtyping

Molecular disease subtype discovery from omics data is an important research problem in precision medicine.The biggest challenges are the skewed distribution and data variability in the measurements of omics data. These challenges complicate the efficient identification of molecular disease subtypes defined by clinical differences, such as survival. Existing approaches adopt kernels to construct patient similarity graphs from each view through pairwise matching. However, the distance functions used in kernels are unable to utilize the potentially critical information of extreme values and data variability which leads to the lack of robustness. In this paper, a novel robust distance metric (ROMDEX) is proposed to construct similarity graphs for molecular disease subtypes from omics data, which is able to address the data variability and extreme values challenges. The proposed approach is validated on multiple TCGA cancer datasets, and the results are compared with multiple baseline disease subtyping methods. The evaluation of results is based on Kaplan-Meier survival time analysis, which is validated using statistical tests e.g, Cox-proportional hazard (Cox p-value). We reject the null hypothesis that the cohorts have the same hazard, for the P-values less than 0.05. The proposed approach achieved best P-values of 0.00181, 0.00171, and 0.00758 for Gene Expression, DNA Methylation, and MicroRNA data respectively, which shows significant difference in survival between the cohorts. In the results, the proposed approach outperformed the existing state-of-the-art (MRGC, PINS, SNF, Consensus Clustering and Icluster+) disease subtyping approaches on various individual disease views of multiple TCGA datasets.

Machine learning-based ABA treatment recommendation and personalization for autism spectrum disorder: an exploratory study

Autism spectrum is a brain development condition that impairs an individual's capacity to communicate socially and manifests through strict routines and obsessive-compulsive behavior. Applied behavior analysis (ABA) is the gold-standard treatment for autism spectrum disorder (ASD). However, as the number of ASD cases increases, there is a substantial shortage of licensed ABA practitioners, limiting the timely formulation, revision, and implementation of treatment plans and goals. Additionally, the subjectivity of the clinician and a lack of data-driven decision-making affect treatment quality. We address these obstacles by applying two machine learning algorithms to recommend and personalize ABA treatment goals for 29 study participants with ASD. The patient similarity and collaborative filtering methods predicted ABA treatment with an average accuracy of 81-84%, with a normalized discounted cumulative gain of 79-81% (NDCG) compared to clinician-prepared ABA treatment recommendations. Additionally, we assess the two models' treatment efficacy (TE) by measuring the percentage of recommended treatment goals mastered by the study participants. The proposed treatment recommendation and personalization strategy are generalizable to other intervention methods in addition to ABA and for other brain disorders. This study was registered as a clinical trial on November 5, 2020 with trial registration number CTRI/2020/11/028933.

Study on the semi-supervised learning-based patient similarity from heterogeneous electronic medical records

Background

A new learning-based patient similarity measurement was proposed to measure patients’ similarity for heterogeneous electronic medical records (EMRs) data.

Methods

We first calculated feature-level similarities according to the features’ attributes. A domain expert provided patient similarity scores of 30 randomly selected patients. These similarity scores and feature-level similarities for 30 patients comprised the labeled sample set, which was used for the semi-supervised learning algorithm to learn the patient-level similarities for all patients. Then we used the k-nearest neighbor (kNN) classifier to predict four liver conditions. The predictive performances were compared in four different situations. We also compared the performances between personalized kNN models and other machine learning models. We assessed the predictive performances by the area under the receiver operating characteristic curve (AUC), F1-score, and cross-entropy (CE) loss.

Results

As the size of the random training samples increased, the kNN models using the learned patient similarity to select near neighbors consistently outperformed those using the Euclidean distance to select near neighbors (all P values < 0.001). The kNN models using the learned patient similarity to identify the top k nearest neighbors from the random training samples also had a higher best-performance (AUC: 0.95 vs. 0.89, F1-score: 0.84 vs. 0.67, and CE loss: 1.22 vs. 1.82) than those using the Euclidean distance. As the size of the similar training samples increased, which composed the most similar samples determined by the learned patient similarity, the performance of kNN models using the simple Euclidean distance to select the near neighbors degraded gradually. When exchanging the role of the Euclidean distance, and the learned patient similarity in selecting the near neighbors and similar training samples, the performance of the kNN models gradually increased. These two kinds of kNN models had the same best-performance of AUC 0.95, F1-score 0.84, and CE loss 1.22. Among the four reference models, the highest AUC and F1-score were 0.94 and 0.80, separately, which were both lower than those for the simple and similarity-based kNN models.

Conclusions

This learning-based method opened an opportunity for similarity measurement based on heterogeneous EMR data and supported the secondary use of EMR data.

Patient similarity analytics for explainable clinical risk prediction

BACKGROUND

Clinical risk prediction models (CRPMs) use patient characteristics to estimate the probability of having or developing a particular disease and/or outcome. While CRPMs are gaining in popularity, they have yet to be widely adopted in clinical practice. The lack of explainability and interpretability has limited their utility. Explainability is the extent of which a model's prediction process can be described. Interpretability is the degree to which a user can understand the predictions made by a model.

METHODS

The study aimed to demonstrate utility of patient similarity analytics in developing an explainable and interpretable CRPM. Data was extracted from the electronic medical records of patients with type-2 diabetes mellitus, hypertension and dyslipidaemia in a Singapore public primary care clinic. We used modified K-nearest neighbour which incorporated expert input, to develop a patient similarity model on this real-world training dataset (n = 7,041) and validated it on a testing dataset (n = 3,018). The results were compared using logistic regression, random forest (RF) and support vector machine (SVM) models from the same dataset. The patient similarity model was then implemented in a prototype system to demonstrate the identification, explainability and interpretability of similar patients and the prediction process.

RESULTS

The patient similarity model (AUROC = 0.718) was comparable to the logistic regression (AUROC = 0.695), RF (AUROC = 0.764) and SVM models (AUROC = 0.766). We packaged the patient similarity model in a prototype web application. A proof of concept demonstrated how the application provided both quantitative and qualitative information, in the form of patient narratives. This information was used to better inform and influence clinical decision-making, such as getting a patient to agree to start insulin therapy.

CONCLUSIONS

Patient similarity analytics is a feasible approach to develop an explainable and interpretable CRPM. While the approach is generalizable, it can be used to develop locally relevant information, based on the database it searches. Ultimately, such an approach can generate a more informative CRPMs which can be deployed as part of clinical decision support tools to better facilitate shared decision-making in clinical practice.

Identification of Similar Patients Through Medical Concept Embedding from Electronic Health Records: A Feasibility Study for Rare Disease Diagnosis

To identify patients with similar clinical profiles and derive insights from the records and outcomes of similar patients can help fast and precise diagnosis and other clinical decisions for rare diseases. Similarity methods are required to take into account the semantic relations between medical concepts and also the different relevance of all medical concepts presented in patients' medical records. In this paper, we introduce the methods developed in the context of rare disease screening/diagnosis from clinical data warehouse using medical concept embedding and adjusted aggregations. Our methods provided better preliminary results than baseline methods, with a significant improvement of precision among the top ranked similar patients, which is encouraging for further fine-tuning and application on a large-scale dataset for new/candidate patient identification.

Methods and measures to quantify ICU patient heterogeneity

Patients in intensive care units are heterogeneous and the daily prediction of their days to discharge (DTD) a complex task that practitioners and computers are not always able to solve satisfactorily. In order to make more precise DTD predictors, it is necessary to have tools for the analysis of the heterogeneity of the patients. Unfortunately, the number of publications in this field is almost non-existent. In order to alleviate this lack of tools, we propose four methods and their corresponding measures to quantify the heterogeneity of intensive patients in the process of determining the DTD. These new methods and measures have been tested with patients admitted over four years to a tertiary hospital in Spain. The results deepen the understanding of the intensive patient and can serve as a basis for the construction of better DTD predictors.

Discovering prescription patterns in pediatric acute-onset neuropsychiatric syndrome patients

OBJECTIVE

Pediatric acute-onset neuropsychiatric syndrome (PANS) is a complex neuropsychiatric syndrome characterized by an abrupt onset of obsessive-compulsive symptoms and/or severe eating restrictions, along with at least two concomitant debilitating cognitive, behavioral, or neurological symptoms. A wide range of pharmacological interventions along with behavioral and environmental modifications, and psychotherapies have been adopted to treat symptoms and underlying etiologies. Our goal was to develop a data-driven approach to identify treatment patterns in this cohort.

MATERIALS AND METHODS

In this cohort study, we extracted medical prescription histories from electronic health records. We developed a modified dynamic programming approach to perform global alignment of those medication histories. Our approach is unique since it considers time gaps in prescription patterns as part of the similarity strategy.

RESULTS

This study included 43 consecutive new-onset pre-pubertal patients who had at least 3 clinic visits. Our algorithm identified six clusters with distinct medication usage history which may represent clinician's practice of treating PANS of different severities and etiologies i.e., two most severe groups requiring high dose intravenous steroids; two arthritic or inflammatory groups requiring prolonged nonsteroidal anti-inflammatory drug (NSAID); and two mild relapsing/remitting group treated with a short course of NSAID. The psychometric scores as outcomes in each cluster generally improved within the first two years.

DISCUSSION AND CONCLUSION

Our algorithm shows potential to improve our knowledge of treatment patterns in the PANS cohort, while helping clinicians understand how patients respond to a combination of drugs.

Diagnosis support systems for rare diseases: a scoping review

INTRODUCTION

Rare diseases affect approximately 350 million people worldwide. Delayed diagnosis is frequent due to lack of knowledge of most clinicians and a small number of expert centers. Consequently, computerized diagnosis support systems have been developed to address these issues, with many relying on rare disease expertise and taking advantage of the increasing volume of generated and accessible health-related data. Our objective is to perform a review of all initiatives aiming to support the diagnosis of rare diseases.

METHODS

A scoping review was conducted based on methods proposed by Arksey and O'Malley. A charting form for relevant study analysis was developed and used to categorize data.

RESULTS

Sixty-eight studies were retained at the end of the charting process. Diagnosis targets varied from 1 rare disease to all rare diseases. Material used for diagnosis support consisted mostly of phenotype concepts, images or fluids. Fifty-seven percent of the studies used expert knowledge. Two-thirds of the studies relied on machine learning algorithms, and one-third used simple similarities. Manual algorithms were encountered as well. Most of the studies presented satisfying performance of evaluation by comparison with references or with external validation. Fourteen studies provided online tools, most of which aimed to support the diagnosis of all rare diseases by considering queries based on phenotype concepts.

CONCLUSION

Numerous solutions relying on different materials and use of various methodologies are emerging with satisfying preliminary results. However, the variability of approaches and evaluation processes complicates the comparison of results. Efforts should be made to adequately validate these tools and guarantee reproducibility and explicability.

A patient-similarity-based model for diagnostic prediction

OBJECTIVE

To simulate the clinical reasoning of doctors, retrieve analogous patients of an index patient automatically and predict diagnoses by the similar/dissimilar patients.

METHODS

We proposed a novel patient-similarity-based framework for diagnostic prediction, which is inspired by the structure-mapping theory about analogy reasoning in psychology. Patient similarity is defined as the similarity between two patients' diagnoses sets rather than a dichotomous (absence/presence of just one disease). The multilabel classification problem is converted to a single-value regression problem by integrating the pairwise patients' clinical features into a vector and taking the vector as the input and the patient similarity as the output. In contrast to the common k-NN method which only considering the nearest neighbors, we not only utilize similar patients (positive analogy) to generate diagnostic hypotheses, but also utilize dissimilar patients (negative analogy) are used to reject diagnostic hypotheses.

RESULTS

The patient-similarity-based models perform better than the one-vs-all baseline and traditional k-NN methods. The f-1 score of positive-analogy-based prediction is 0.698, significantly higher than the scores of baselines ranging from 0.368 to 0.661. It increases to 0.703 when the negative analogy method is applied to modify the prediction results of positive analogy. The performance of this method is highly promising for larger datasets.

CONCLUSION

The patient-similarity-based model provides diagnostic decision support that is more accurate, generalizable, and interpretable than those of previous methods and is based on heterogeneous and incomplete data. The model also serves as a new application for the use of clinical big data through artificial intelligence technology.

Evaluating global and local sequence alignment methods for comparing patient medical records

Background

Sequence alignment is a way of arranging sequences (e.g., DNA, RNA, protein, natural language, financial data, or medical events) to identify the relatedness between two or more sequences and regions of similarity. For Electronic Health Records (EHR) data, sequence alignment helps to identify patients of similar disease trajectory for more relevant and precise prognosis, diagnosis and treatment of patients.

Methods

We tested two cutting-edge global sequence alignment methods, namely dynamic time warping (DTW) and Needleman-Wunsch algorithm (NWA), together with their local modifications, DTW for Local alignment (DTWL) and Smith-Waterman algorithm (SWA), for aligning patient medical records. We also used 4 sets of synthetic patient medical records generated from a large real-world EHR database as gold standard data, to objectively evaluate these sequence alignment algorithms.

Results

For global sequence alignments, 47 out of 80 DTW alignments and 11 out of 80 NWA alignments had superior similarity scores than reference alignments while the rest 33 DTW alignments and 69 NWA alignments had the same similarity scores as reference alignments. Forty-six out of 80 DTW alignments had better similarity scores than NWA alignments with the rest 34 cases having the equal similarity scores from both algorithms. For local sequence alignments, 70 out of 80 DTWL alignments and 68 out of 80 SWA alignments had larger coverage and higher similarity scores than reference alignments while the rest DTWL alignments and SWA alignments received the same coverage and similarity scores as reference alignments. Six out of 80 DTWL alignments showed larger coverage and higher similarity scores than SWA alignments. Thirty DTWL alignments had the equal coverage but better similarity scores than SWA. DTWL and SWA received the equal coverage and similarity scores for the rest 44 cases.

Conclusions

DTW, NWA, DTWL and SWA outperformed the reference alignments. DTW (or DTWL) seems to align better than NWA (or SWA) by inserting new daily events and identifying more similarities between patient medical records. The evaluation results could provide valuable information on the strengths and weakness of these sequence alignment methods for future development of sequence alignment methods and patient similarity-based studies.

Measurement and application of patient similarity in personalized predictive modeling based on electronic medical records

Background

Conventional risk prediction techniques may not be the most suitable approach for personalized prediction for individual patients. Therefore, individualized predictive modeling based on similar patients has emerged. This study aimed to propose a comprehensive measurement of patient similarity using real-world electronic medical records data, and evaluate the effectiveness of the individualized prediction of a patient’s diabetes status based on the patient similarity.

Results

When using no more than 30% of the whole training sample, the personalized predictive models outperformed corresponding traditional models built on randomly selected training samples of the same size as the personalized models (P < 0.001 for all). With only the top 1000 (10%), 700 (7%) and 1400 (14%) similar samples, personalized random forest, k-nearest neighbor and logistic regression models reached the globally optimal performance with the area under the receiver-operating characteristic (ROC) curve of 0.90, 0.82 and 0.89, respectively.

Conclusions

The proposed patient similarity measurement was effective when developing personalized predictive models. The successful application of patient similarity in predicting a patient’s diabetes status provided useful references for diagnostic decision-making support by investigating the evidence on similar patients.

Using the distance between sets of hierarchical taxonomic clinical concepts to measure patient similarity

Background

Many clinical concepts are standardized under a categorical and hierarchical taxonomy such as ICD-10, ATC, etc. These taxonomic clinical concepts provide insight into semantic meaning and similarity among clinical concepts and have been applied to patient similarity measures. However, the effects of diverse set sizes of taxonomic clinical concepts contributing to similarity at the patient level have not been well studied.

Methods

In this paper the most widely used taxonomic clinical concepts system, ICD-10, was studied as a representative taxonomy. The distance between ICD-10-coded diagnosis sets is an integrated estimation of the information content of each concept, the similarity between each pairwise concepts and the similarity between the sets of concepts. We proposed a novel method at the set-level similarity to calculate the distance between sets of hierarchical taxonomic clinical concepts to measure patient similarity. A real-world clinical dataset with ICD-10 coded diagnoses and hospital length of stay (HLOS) information was used to evaluate the performance of various algorithms and their combinations in predicting whether a patient need long-term hospitalization or not. Four subpopulation prototypes that were defined based on age and HLOS with different diagnoses set sizes were used as the target for similarity analysis. The F-score was used to evaluate the performance of different algorithms by controlling other factors. We also evaluated the effect of prototype set size on prediction precision.

Results

The results identified the strengths and weaknesses of different algorithms to compute information content, code-level similarity and set-level similarity under different contexts, such as set size and concept set background. The minimum weighted bipartite matching approach, which has not been fully recognized previously showed unique advantages in measuring the concepts-based patient similarity.

Conclusions

This study provides a systematic benchmark evaluation of previous algorithms and novel algorithms used in taxonomic concepts-based patient similarity, and it provides the basis for selecting appropriate methods under different clinical scenarios.

Electronic supplementary material

The online version of this article (10.1186/s12911-019-0807-y) contains supplementary material, which is available to authorized users.

Heart Disease Dataset Clusterization

Clusterization is a promising group of methods in the context of patient similarity. However, results of clustering are not often clear for physicians as well as different clustering methods can produce different results. We have examined a well-known dataset and implemented 3 clustering methods (k-means, Agglomerative and Spectral). We have compared and evaluated clusters and their correlation with data attributes. In contrast to original dataset's target value, the clusters correlated with only a few attributes. Finally, we train 2 predictive models based on k-nearest neighbors (KNN) algorithm and Artificial Neural Network (ANN). Models evaluation demonstrates that using the results of clustering algorithms as predictive attribute give a higher F-score than the original target attribute.

Patient similarity for precision medicine: A systematic review

Evidence-based medicine is the most prevalent paradigm adopted by physicians. Clinical practice guidelines typically define a set of recommendations together with eligibility criteria that restrict their applicability to a specific group of patients. The ever-growing size and availability of health-related data is currently challenging the broad definitions of guideline-defined patient groups. Precision medicine leverages on genetic, phenotypic, or psychosocial characteristics to provide precise identification of patient subsets for treatment targeting. Defining a patient similarity measure is thus an essential step to allow stratification of patients into clinically-meaningful subgroups. The present review investigates the use of patient similarity as a tool to enable precision medicine. 279 articles were analyzed along four dimensions: data types considered, clinical domains of application, data analysis methods, and translational stage of findings. Cancer-related research employing molecular profiling and standard data analysis techniques such as clustering constitute the majority of the retrieved studies. Chronic and psychiatric diseases follow as the second most represented clinical domains. Interestingly, almost one quarter of the studies analyzed presented a novel methodology, with the most advanced employing data integration strategies and being portable to different clinical domains. Integration of such techniques into decision support systems constitutes and interesting trend for future research.

A Predictive Model for Patient Similarity: Classes Based on Secondary Data and Simple Measurements as Predictors

Predictive models optimized for average cases might work not perfect for cases deviating from average because they are based on a cohort of all patients. Models could be more personalized if they were built on a sub-cohort of patients similar to a current one and to train models on data collected from those similar patients. In this paper, we consider patient similarity as a classification task. We suppose that data such as diagnoses and treatment obtained by physicians (secondary data) are more relevant for similarity than tests and measurements (primary data). We defined several classes based on diagnoses and outcomes and apply a predictive model for classification. We used five commonly used and easy to obtain measurements as predictors for the model. All measurements were collected during the first 24 hours after admission. We have shown that classes of similar patients can be defined on the basis of a previous patient's secondary data and new patients can be classified into these classes.

A Novel Temporal Similarity Measure for Patients Based on Irregularly Measured Data in Electronic Health Records

Patient similarity measurement is an important tool for cohort identification in clinical decision support applications. A reliable similarity metric can be used for deriving diagnostic or prognostic information about a target patient using other patients with similar trajectories of health-care events. However, the measure of similar care trajectories is challenged by the irregularity of measurements, inherent in health care. To address this challenge, we propose a novel temporal similarity measure for patients based on irregularly measured laboratory test data from the Multiparameter Intelligent Monitoring in Intensive Care database and the pediatric Intensive Care Unit (ICU) database of Children's Healthcare of Atlanta. This similarity measure, which is modified from the Smith Waterman algorithm, identifies patients that share sequentially similar laboratory results separated by time intervals of similar length. We demonstrate the predictive power of our method; that is, patients with higher similarity in their previous histories will most likely have higher similarity in their later histories. In addition, compared with other non-temporal measures, our method is stronger at predicting mortality in ICU patients diagnosed with acute kidney injury and sepsis.

Categories and Subject Descriptors

H.3.3 [Information Storage and Retrieval]: Retrieval models and rankings - similarity measures; J.3 [Applied Computing]: Life and medical sciences - health and medical information systems.

General Term

Algorithm.