Automatic classification of scanned electronic health record documents.
Ontology highlight
ABSTRACT: Objectives
Electronic Health Records (EHRs) contain scanned documents from a variety of sources such as identification cards, radiology reports, clinical correspondence, and many other document types. We describe the distribution of scanned documents at one health institution and describe the design and evaluation of a system to categorize documents into clinically relevant and non-clinically relevant categories as well as further sub-classifications. Our objective is to demonstrate that text classification systems can accurately classify scanned documents.Methods
We extracted text using Optical Character Recognition (OCR). We then created and evaluated multiple text classification machine learning models, including both "bag of words" and deep learning approaches. We evaluated the system on three different levels of classification using both the entire document as input, as well as the individual pages of the document. Finally, we compared the effects of different text processing methods.Results
A deep learning model using ClinicalBERT performed best. This model distinguished between clinically-relevant documents and not clinically-relevant documents with an accuracy of 0.973; between intermediate sub-classifications with an accuracy of 0.949; and between individual classes with an accuracy of 0.913.Discussion
Within the EHR, some document categories such as "external medical records" may contain hundreds of scanned pages without clear document boundaries. Without further sub-classification, clinicians must view every page or risk missing clinically-relevant information. Machine learning can automatically classify these scanned documents to reduce clinician burden.Conclusion
Using machine learning applied to OCR-extracted text has the potential to accurately identify clinically-relevant scanned content within EHRs.
SUBMITTER: Goodrum H
PROVIDER: S-EPMC7731898 | biostudies-literature |
REPOSITORIES: biostudies-literature
Similar Datasets
Project description:IntroductionElectronic health records (EHR) are linked together to examine disease history and to undertake research into the causes and outcomes of disease. However, the process of constructing algorithms for phenotyping (e.g., identifying disease characteristics) or health characteristics (e.g., smoker) is very time consuming and resource costly. In addition, results can vary greatly between researchers. Reusing or building on algorithms that others have created is a compelling solution to these problems. However, sharing algorithms is not a common practice and many published studies do not detail the clinical code lists used by the researchers in the disease/characteristic definition. To address these challenges, a number of centres across the world have developed health data portals which contain concept libraries (e.g., algorithms for defining concepts such as disease and characteristics) in order to facilitate disease phenotyping and health studies.ObjectivesThis study aims to review the literature of existing concept libraries, examine their utilities, identify the current gaps, and suggest future developments.MethodsThe five-stage framework of Arksey and O'Malley was used for the literature search. This approach included defining the research questions, identifying relevant studies through literature review, selecting eligible studies, charting and extracting data, and summarising and reporting the findings.ResultsThis review identified seven publicly accessible Electronic Health data concept libraries which were developed in different countries including UK, USA, and Canada. The concept libraries (n = 7) investigated were either general libraries that hold phenotypes of multiple specialties (n = 4) or specialized libraries that manage only certain specialities such as rare diseases (n = 3). There were some clear differences between the general libraries such as archiving data from different electronic sources, and using a range of different types of coding systems. However, they share some clear similarities such as enabling users to upload their own code lists, and allowing users to use/download the publicly accessible code. In addition, there were some differences between the specialized libraries such as difference in ability to search, and if it was possible to use different searching queries such as simple or complex searches. Conversely, there were some similarities between the specialized libraries such as enabling users to upload their own concepts into the libraries and to show where they were published, which facilitates assessing the validity of the concepts. All the specialized libraries aimed to encourage the reuse of research methods such as lists of clinical code and/or metadata.ConclusionThe seven libraries identified have been developed independently and appear to replicate similar concepts but in different ways. Collaboration between similar libraries would greatly facilitate the use of these libraries for the user. The process of building code lists takes time and effort. Access to existing code lists increases consistency and accuracy of definitions across studies. Concept library developers should collaborate with each other to raise awareness of their existence and of their various functions, which could increase users' contributions to those libraries and promote their wide-ranging adoption.
| S-EPMC8210840 | biostudies-literature
Project description:Background and study aims
The Federal Ministry of Education and Research (BMBF) has funded the development of a patient controlled “Personal Electronic Health Record” (PEPA) as part of their current INFOPAT project. PEPA allows for the exchange of medical data between all interested groups, including the patients thenselves. A PEPA contains medical reports and findings including, for example, x-rays, CT scans and MRIs. The patients can safeguard their information and access all content. The aim of this study is to test the PEPA-prototype for cancer patients being treated at the National Center for Tumor Diseases (NCT) in Heidelberg.
Who can participate?
Adult colorectal cancer patients being treated at the National Center for Tumor Diseases (NCT) in Heidelberg.
What does the study involve?
There are two part to this study. The first part involves focus groups, in which a group of people are asked their opinion and attitudes towards, for example, an idea, product or service. Here, focus groups are set up that involve the patients, doctors and other health professionals to discuss current problems with the coordination of care offered to cancer patients, as well as discuss what people want to get from using PEPA (i.e. user requirements). The PEPA-prototype is then developed taking into account the conclusions from the focus groups. In the second part of the study, patients are encouraged to use the PEPA-prototype to prepare and follow up meetings with their physicians.
| 2424890 | ecrin-mdr-crc
Project description:BACKGROUND:Computable phenotypes have the ability to utilize data within the electronic health record (EHR) to identify patients with certain characteristics. Many computable phenotypes rely on multiple types of data within the EHR including prescription drug information. Hypertension (HTN)-related computable phenotypes are particularly dependent on the correct classification of antihypertensive prescription drug information, as well as corresponding diagnoses and blood pressure information. OBJECTIVE:This study aimed to create an antihypertensive drug classification system to be utilized with EHR-based data as part of HTN-related computable phenotypes. METHODS:We compared 4 different antihypertensive drug classification systems based off of 4 different methodologies and terminologies, including 3 RxNorm Concept Unique Identifier (RxCUI)-based classifications and 1 medication name-based classification. The RxCUI-based classifications utilized data from (1) the Drug Ontology, (2) the new Medication Reference Terminology, and (3) the Anatomical Therapeutic Chemical Classification System and DrugBank, whereas the medication name-based classification relied on antihypertensive drug names. Each classification system was applied to EHR-based prescription drug data from hypertensive patients in the OneFlorida Data Trust. RESULTS:There were 13,627 unique RxCUIs and 8025 unique medication names from the 13,879,046 prescriptions. We observed a broad overlap between the 4 methods, with 84.1% (691/822) to 95.3% (695/729) of terms overlapping pairwise between the different classification methods. Key differences arose from drug products with multiple dosage forms, drug products with an indication of benign prostatic hyperplasia, drug products that contain more than 1 ingredient (combination products), and terms within the classification systems corresponding to retired or obsolete RxCUIs. CONCLUSIONS:In total, 2 antihypertensive drug classifications were constructed, one based on RxCUIs and one based on medication name, that can be used in future computable phenotypes that require antihypertensive drug classifications.
| S-EPMC7068459 | biostudies-literature
Project description:BackgroundIt is assumed that the implementation of health information technology introduces new vulnerabilities within a complex sociotechnical health care system, but no international consensus exists on a standardized format for enhancing the collection, analysis, and interpretation of technology-induced errors.ObjectiveThis study aims to develop a classification for patient safety incident reporting associated with the use of mature electronic health records (EHRs). It also aims to validate the classification by using a data set of incidents during a 6-month period immediately after the implementation of a new EHR system.MethodsThe starting point of the classification development was the Finnish Technology-Induced Error Risk Assessment Scale tool, based on research on commonly recognized error types. A multiprofessional research team used iterative tests on consensus building to develop a classification system. The final classification, with preliminary descriptions of classes, was validated by applying it to analyze EHR-related error incidents (n=428) during the implementation phase of a new EHR system and also to evaluate this classification's characteristics and applicability for reporting incidents. Interrater agreement was applied.ResultsThe number of EHR-related patient safety incidents during the implementation period (n=501) was five-fold when compared with the preimplementation period (n=82). The literature identified new error types that were added to the emerging classification. Error types were adapted iteratively after several test rounds to develop a classification for reporting patient safety incidents in the clinical use of a high-maturity EHR system. Of the 427 classified patient safety incidents, interface problems accounted for 96 (22.5%) incident reports, usability problems for 73 (17.1%), documentation problems for 60 (14.1%), and clinical workflow problems for 33 (7.7%). Altogether, 20.8% (89/427) of reports were related to medication section problems, and downtime problems were rare (n=8). During the classification work, 14.8% (74/501) of reports of the original sample were rejected because of insufficient information, even though the reports were deemed to be related to EHRs. The interrater agreement during the blinded review was 97.7%.ConclusionsThis study presents a new classification for EHR-related patient safety incidents applicable to mature EHRs. The number of EHR-related patient safety incidents during the implementation period may reflect patient safety challenges during the implementation of a new type of high-maturity EHR system. The results indicate that the types of errors previously identified in the literature change with the EHR development cycle.
| S-EPMC8441612 | biostudies-literature
Project description:ObjectivesSocial determinants of health (SDH), key contributors to health, are rarely systematically measured and collected in the electronic health record (EHR). We investigate how to leverage clinical notes using novel applications of multi-label learning (MLL) to classify SDH in mental health and substance use disorder patients who frequent the emergency department.Methods and materialsWe labeled a gold-standard corpus of EHR clinical note sentences (N = 4063) with 6 identified SDH-related domains recommended by the Institute of Medicine for inclusion in the EHR. We then trained 5 classification models: linear-Support Vector Machine, K-Nearest Neighbors, Random Forest, XGBoost, and bidirectional Long Short-Term Memory (BI-LSTM). We adopted 5 common evaluation measures: accuracy, average precision-recall (AP), area under the curve receiver operating characteristic (AUC-ROC), Hamming loss, and log loss to compare the performance of different methods for MLL classification using the F1 score as the primary evaluation metric.ResultsOur results suggested that, overall, BI-LSTM outperformed the other classification models in terms of AUC-ROC (93.9), AP (0.76), and Hamming loss (0.12). The AUC-ROC values of MLL models of SDH related domains varied between (0.59-1.0). We found that 44.6% of our study population (N = 1119) had at least one positive documentation of SDH.Discussion and conclusionThe proposed approach of training an MLL model on an SDH rich data source can produce a high performing classifier using only unstructured clinical notes. We also provide evidence that model performance is associated with lexical diversity by health professionals and the auto-generation of clinical note sentences to document SDH.
| S-EPMC8423426 | biostudies-literature
Project description:BACKGROUND:Hypoglycemic events are common and potentially dangerous conditions among patients being treated for diabetes. Automatic detection of such events could improve patient care and is valuable in population studies. Electronic health records (EHRs) are valuable resources for the detection of such events. OBJECTIVE:In this study, we aim to develop a deep-learning-based natural language processing (NLP) system to automatically detect hypoglycemic events from EHR notes. Our model is called the High-Performing System for Automatically Detecting Hypoglycemic Events (HYPE). METHODS:Domain experts reviewed 500 EHR notes of diabetes patients to determine whether each sentence contained a hypoglycemic event or not. We used this annotated corpus to train and evaluate HYPE, the high-performance NLP system for hypoglycemia detection. We built and evaluated both a classical machine learning model (ie, support vector machines [SVMs]) and state-of-the-art neural network models. RESULTS:We found that neural network models outperformed the SVM model. The convolutional neural network (CNN) model yielded the highest performance in a 10-fold cross-validation setting: mean precision=0.96 (SD 0.03), mean recall=0.86 (SD 0.03), and mean F1=0.91 (SD 0.03). CONCLUSIONS:Despite the challenges posed by small and highly imbalanced data, our CNN-based HYPE system still achieved a high performance for hypoglycemia detection. HYPE can be used for EHR-based hypoglycemia surveillance and population studies in diabetes patients.
| S-EPMC6913754 | biostudies-literature
Project description:BACKGROUND:Accurate anesthesiology procedure code data are essential to quality improvement, research, and reimbursement tasks within anesthesiology practices. Advanced data science techniques, including machine learning and natural language processing, offer opportunities to develop classification tools for Current Procedural Terminology codes across anesthesia procedures. METHODS:Models were created using a Train/Test dataset including 1,164,343 procedures from 16 academic and private hospitals. Five supervised machine learning models were created to classify anesthesiology Current Procedural Terminology codes, with accuracy defined as first choice classification matching the institutional-assigned code existing in the perioperative database. The two best performing models were further refined and tested on a Holdout dataset from a single institution distinct from Train/Test. A tunable confidence parameter was created to identify cases for which models were highly accurate, with the goal of at least 95% accuracy, above the reported 2018 Centers for Medicare and Medicaid Services (Baltimore, Maryland) fee-for-service accuracy. Actual submitted claim data from billing specialists were used as a reference standard. RESULTS:Support vector machine and neural network label-embedding attentive models were the best performing models, respectively, demonstrating overall accuracies of 87.9% and 84.2% (single best code), and 96.8% and 94.0% (within top three). Classification accuracy was 96.4% in 47.0% of cases using support vector machine and 94.4% in 62.2% of cases using label-embedding attentive model within the Train/Test dataset. In the Holdout dataset, respective classification accuracies were 93.1% in 58.0% of cases and 95.0% among 62.0%. The most important feature in model training was procedure text. CONCLUSIONS:Through application of machine learning and natural language processing techniques, highly accurate real-time models were created for anesthesiology Current Procedural Terminology code classification. The increased processing speed and a priori targeted accuracy of this classification approach may provide performance optimization and cost reduction for quality improvement, research, and reimbursement tasks reliant on anesthesiology procedure codes.
| S-EPMC7665375 | biostudies-literature
Project description:ObjectivesScanned documents (SDs), while common in electronic health records and potentially rich in clinically relevant information, rarely fit well with clinician workflow. Here, we identify scanned imaging reports requiring follow-up with high recall and practically useful precision.Materials and methodsWe focused on identifying imaging findings for 3 common causes of malpractice claims: (1) potentially malignant breast (mammography) and (2) lung (chest computed tomography [CT]) lesions and (3) long-bone fracture (X-ray) reports. We train our ClinicalBERT-based pipeline on existing typed/dictated reports classified manually or using ICD-10 codes, evaluate using a test set of manually classified SDs, and compare against string-matching (baseline approach).ResultsA total of 393 mammograms, 305 chest CT, and 683 bone X-ray reports were manually reviewed. The string-matching approach had an F1 of 0.667. For mammograms, chest CTs, and bone X-rays, respectively: models trained on manually classified training data and optimized for F1 reached an F1 of 0.900, 0.905, and 0.817, while separate models optimized for recall achieved a recall of 1.000 with precisions of 0.727, 0.518, and 0.275. Models trained on ICD-10-labelled data and optimized for F1 achieved F1 scores of 0.647, 0.830, and 0.643, while those optimized for recall achieved a recall of 1.0 with precisions of 0.407, 0.683, and 0.358.DiscussionOur pipeline can identify abnormal reports with potentially useful performance and so decrease the manual effort required to screen for abnormal findings that require follow-up.ConclusionIt is possible to automatically identify clinically significant abnormalities in SDs with high recall and practically useful precision in a generalizable and minimally laborious way.
| S-EPMC9714594 | biostudies-literature
Project description:BackgroundThe rapid expansion in the use of electronic health records (EHR) has increased the number of medical errors originating in health information systems (HIS). The sociotechnical approach helps in understanding risks in the development, implementation, and use of EHR and health information technology (HIT) while accounting for complex interactions of technology within the health care system.ObjectiveThis study addresses two important questions: (1) "which of the common EHR error types are associated with perceived high- and extreme-risk severity ratings among EHR users?", and (2) "which variables are associated with high- and extreme-risk severity ratings?"MethodsThis study was a quantitative, non-experimental, descriptive study of EHR users. We conducted a cross-sectional web-based questionnaire study at the largest hospital district in Finland. Statistical tests included the reliability of the summative scales tested with Cronbach's alpha. Logistic regression served to assess the association of the independent variables to each of the eight risk factors examined.ResultsA total of 2864 eligible respondents provided the final data. Almost half of the respondents reported a high level of risk related to the error type "extended EHR unavailability". The lowest overall risk level was associated with "selecting incorrectly from a list of items". In multivariate analyses, profession and clinical unit proved to be the strongest predictors for high perceived risk. Physicians perceived risk levels to be the highest ( P<.001 in six of eight error types), while emergency departments, operating rooms, and procedure units were associated with higher perceived risk levels ( P<.001 in four of eight error types). Previous participation in eLearning courses on EHR-use was associated with lower risk for some of the risk factors.ConclusionsBased on a large number of Finnish EHR users in hospitals, this study indicates that HIT safety hazards should be taken very seriously, particularly in operating rooms, procedure units, emergency departments, and intensive care units/critical care units. Health care organizations should use proactive and systematic assessments of EHR risks before harmful events occur. An EHR training program should be compulsory for all EHR users in order to address EHR safety concerns resulting from the failure to use HIT appropriately.
| S-EPMC4890731 | biostudies-literature
Project description:ObjectiveExcess physician work hours contribute to burnout and medical errors. Self-report of work hours is burdensome and often inaccurate. We aimed to validate a method that automatically determines provider shift duration based on electronic health record (EHR) timestamps across multiple inpatient settings within a single institution.MethodsWe developed an algorithm to calculate shift start and end times for inpatient providers based on EHR timestamps. We validated the algorithm based on overlap between calculated shifts and scheduled shifts. We then demonstrated a use case by calculating shifts for pediatric residents on inpatient rotations from July 1, 2015 through June 30, 2016, comparing hours worked and number of shifts by rotation and role.ResultsWe collected 6.3 × 107 EHR timestamps for 144 residents on 771 inpatient rotations, yielding 14,678 EHR-calculated shifts. Validation on a subset of shifts demonstrated 100% shift match and 87.9 ± 0.3% overlap (mean ± standard error [SE]) with scheduled shifts. Senior residents functioning as front-line clinicians worked more hours per 4-week block (mean ± SE: 273.5 ± 1.7) than senior residents in supervisory roles (253 ± 2.3) and junior residents (241 ± 2.5). Junior residents worked more shifts per block (21 ± 0.1) than senior residents (18 ± 0.1).ConclusionAutomatic calculation of inpatient provider work hours is feasible using EHR timestamps. An algorithm to assess provider work hours demonstrated criterion validity via comparison with scheduled shifts. Differences between junior and senior residents in calculated mean hours worked and number of shifts per 4-week block were also consistent with differences in scheduled shifts and duty-hour restrictions.
| S-EPMC8438622 | biostudies-literature