Dataset Information

Performance of a Machine Learning Algorithm Using Electronic Health Record Data to Identify and Estimate Survival in a Longitudinal Cohort of Patients With Lung Cancer.

ABSTRACT:

Importance

Electronic health records (EHRs) provide a low-cost means of accessing detailed longitudinal clinical data for large populations. A lung cancer cohort assembled from EHR data would be a powerful platform for clinical outcome studies.

Objective

To investigate whether a clinical cohort assembled from EHRs could be used in a lung cancer prognosis study.

Design, setting, and participants

In this cohort study, patients with lung cancer were identified among 76 643 patients with at least 1 lung cancer diagnostic code deposited in an EHR in Mass General Brigham health care system from July 1988 to October 2018. Patients were identified via a semisupervised machine learning algorithm, for which clinical information was extracted from structured and unstructured data via natural language processing tools. Data completeness and accuracy were assessed by comparing with the Boston Lung Cancer Study and against criterion standard EHR review results. A prognostic model for non-small cell lung cancer (NSCLC) overall survival was further developed for clinical application. Data were analyzed from March 2019 through July 2020.

Exposures

Clinical data deposited in EHRs for cohort construction and variables of interest for the prognostic model were collected.

Main outcomes and measures

The primary outcomes were the performance of the lung cancer classification model and the quality of the extracted variables; the secondary outcome was the performance of the prognostic model.

Results

Among 76 643 patients with at least 1 lung cancer diagnostic code, 42 069 patients were identified as having lung cancer, with a positive predictive value of 94.4%. The study cohort consisted of 35 375 patients (16 613 men [47.0%] and 18 756 women [53.0%]; 30 140 White individuals [85.2%], 1040 Black individuals [2.9%], and 857 Asian individuals [2.4%]) after excluding patients with lung cancer history and less than 14 days of follow-up after initial diagnosis. The median (interquartile range) age at diagnosis was 66.7 (58.4-74.1) years. The area under the receiver operating characteristic curves of the prognostic model for overall survival with NSCLC were 0.828 (95% CI, 0.815-0.842) for 1-year prediction, 0.825 (95% CI, 0.812-0.836) for 2-year prediction, 0.814 (95% CI, 0.800-0.826) for 3-year prediction, 0.814 (95% CI, 0.799-0.828) for 4-year prediction, and 0.812 (95% CI, 0.798-0.825) for 5-year prediction.

Conclusions and relevance

These findings suggest the feasibility of assembling a large-scale EHR-based lung cancer cohort with detailed longitudinal clinical measurements and that EHR data may be applied in cancer progression with a set of generalizable approaches.

SUBMITTER: Yuan Q

PROVIDER: S-EPMC8264641 | biostudies-literature |

REPOSITORIES: biostudies-literature

ACCESS DATA

Similar Datasets

Project description:BackgroundA major problem in treating acute kidney injury (AKI) is that clinical criteria for recognition are markers of established kidney damage or impaired function; treatment before such damage manifests is desirable. Clinicians could intervene during what may be a crucial stage for preventing permanent kidney injury if patients with incipient AKI and those at high risk of developing AKI could be identified.ObjectiveIn this study, we evaluate a machine learning algorithm for early detection and prediction of AKI.DesignWe used a machine learning technique, boosted ensembles of decision trees, to train an AKI prediction tool on retrospective data taken from more than 300 000 inpatient encounters.SettingData were collected from inpatient wards at Stanford Medical Center and intensive care unit patients at Beth Israel Deaconess Medical Center.PatientsPatients older than the age of 18 whose hospital stays lasted between 5 and 1000 hours and who had at least one documented measurement of heart rate, respiratory rate, temperature, serum creatinine (SCr), and Glasgow Coma Scale (GCS).MeasurementsWe tested the algorithm's ability to detect AKI at onset and to predict AKI 12, 24, 48, and 72 hours before onset.MethodsWe tested AKI detection and prediction using the National Health Service (NHS) England AKI Algorithm as a gold standard. We additionally tested the algorithm's ability to detect AKI as defined by the Kidney Disease: Improving Global Outcomes (KDIGO) guidelines. We compared the algorithm's 3-fold cross-validation performance to the Sequential Organ Failure Assessment (SOFA) score for AKI identification in terms of area under the receiver operating characteristic (AUROC).ResultsThe algorithm demonstrated high AUROC for detecting and predicting NHS-defined AKI at all tested time points. The algorithm achieves AUROC of 0.872 (95% confidence interval [CI], 0.867-0.878) for AKI detection at time of onset. For prediction 12 hours before onset, the algorithm achieves an AUROC of 0.800 (95% CI, 0.792-0.809). For 24-hour predictions, the algorithm achieves AUROC of 0.795 (95% CI, 0.785-0.804). For 48-hour and 72-hour predictions, the algorithm achieves AUROC values of 0.761 (95% CI, 0.753-0.768) and 0.728 (95% CI, 0.719-0.737), respectively.LimitationsBecause of the retrospective nature of this study, we cannot draw any conclusions about the impact the algorithm's predictions will have on patient outcomes in a clinical setting.ConclusionsThe results of these experiments suggest that a machine learning-based AKI prediction tool may offer important prognostic capabilities for determining which patients are likely to suffer AKI, potentially allowing clinicians to intervene before kidney damage manifests.

Project description:OBJECTIVES:The goal of this study was to use machine learning to more accurately predict survival after echocardiography. BACKGROUND:Predicting patient outcomes (e.g., survival) following echocardiography is primarily based on ejection fraction (EF) and comorbidities. However, there may be significant predictive information within additional echocardiography-derived measurements combined with clinical electronic health record data. METHODS:Mortality was studied in 171,510 unselected patients who underwent 331,317 echocardiograms in a large regional health system. The authors investigated the predictive performance of nonlinear machine learning models compared with that of linear logistic regression models using 3 different inputs: 1) clinical variables, including 90 cardiovascular-relevant International Classification of Diseases, Tenth Revision, codes, and age, sex, height, weight, heart rate, blood pressures, low-density lipoprotein, high-density lipoprotein, and smoking; 2) clinical variables plus physician-reported EF; and 3) clinical variables and EF, plus 57 additional echocardiographic measurements. Missing data were imputed with a multivariate imputation by using a chained equations algorithm (MICE). The authors compared models versus each other and baseline clinical scoring systems by using a mean area under the curve (AUC) over 10 cross-validation folds and across 10 survival durations (6 to 60 months). RESULTS:Machine learning models achieved significantly higher prediction accuracy (all AUC >0.82) over common clinical risk scores (AUC = 0.61 to 0.79), with the nonlinear random forest models outperforming logistic regression (p < 0.01). The random forest model including all echocardiographic measurements yielded the highest prediction accuracy (p < 0.01 across all models and survival durations). Only 10 variables were needed to achieve 96% of the maximum prediction accuracy, with 6 of these variables being derived from echocardiography. Tricuspid regurgitation velocity was more predictive of survival than LVEF. In a subset of studies with complete data for the top 10 variables, multivariate imputation by chained equations yielded slightly reduced predictive accuracies (difference in AUC of 0.003) compared with the original data. CONCLUSIONS:Machine learning can fully utilize large combinations of disparate input variables to predict survival after echocardiography with superior accuracy.

Project description:BackgroundFinancial codes are often used to extract diagnoses from electronic health records. This approach is prone to false positives. Alternatively, queries are constructed, but these are highly center and language specific. A tantalizing alternative is the automatic identification of patients by employing machine learning on format-free text entries.ObjectiveThe aim of this study was to develop an easily implementable workflow that builds a machine learning algorithm capable of accurately identifying patients with rheumatoid arthritis from format-free text fields in electronic health records.MethodsTwo electronic health record data sets were employed: Leiden (n=3000) and Erlangen (n=4771). Using a portion of the Leiden data (n=2000), we compared 6 different machine learning methods and a naïve word-matching algorithm using 10-fold cross-validation. Performances were compared using the area under the receiver operating characteristic curve (AUROC) and the area under the precision recall curve (AUPRC), and F1 score was used as the primary criterion for selecting the best method to build a classifying algorithm. We selected the optimal threshold of positive predictive value for case identification based on the output of the best method in the training data. This validation workflow was subsequently applied to a portion of the Erlangen data (n=4293). For testing, the best performing methods were applied to remaining data (Leiden n=1000; Erlangen n=478) for an unbiased evaluation.ResultsFor the Leiden data set, the word-matching algorithm demonstrated mixed performance (AUROC 0.90; AUPRC 0.33; F1 score 0.55), and 4 methods significantly outperformed word-matching, with support vector machines performing best (AUROC 0.98; AUPRC 0.88; F1 score 0.83). Applying this support vector machine classifier to the test data resulted in a similarly high performance (F1 score 0.81; positive predictive value [PPV] 0.94), and with this method, we could identify 2873 patients with rheumatoid arthritis in less than 7 seconds out of the complete collection of 23,300 patients in the Leiden electronic health record system. For the Erlangen data set, gradient boosting performed best (AUROC 0.94; AUPRC 0.85; F1 score 0.82) in the training set, and applied to the test data, resulted once again in good results (F1 score 0.67; PPV 0.97).ConclusionsWe demonstrate that machine learning methods can extract the records of patients with rheumatoid arthritis from electronic health record data with high precision, allowing research on very large populations for limited costs. Our approach is language and center independent and could be applied to any type of diagnosis. We have developed our pipeline into a universally applicable and easy-to-implement workflow to equip centers with their own high-performing algorithm. This allows the creation of observational studies of unprecedented size covering different countries for low cost from already available data in electronic health record systems.

Project description:BACKGROUND:Systemic sclerosis (SSc) is a rare disease with studies limited by small sample sizes. Electronic health records (EHRs) represent a powerful tool to study patients with rare diseases such as SSc, but validated methods are needed. We developed and validated EHR-based algorithms that incorporate billing codes and clinical data to identify SSc patients in the EHR. METHODS:We used a de-identified EHR with over 3 million subjects and identified 1899 potential SSc subjects with at least 1 count of the SSc ICD-9 (710.1) or ICD-10-CM (M34*) codes. We randomly selected 200 as a training set for chart review. A subject was a case if diagnosed with SSc by a rheumatologist, dermatologist, or pulmonologist. We selected the following algorithm components based on clinical knowledge and available data: SSc ICD-9 and ICD-10-CM codes, positive antinuclear antibody (ANA) (titer ??1:80), and a keyword of Raynaud's phenomenon (RP). We performed both rule-based and machine learning techniques for algorithm development. Positive predictive values (PPVs), sensitivities, and F-scores (which account for PPVs and sensitivities) were calculated for the algorithms. RESULTS:PPVs were low for algorithms using only 1 count of the SSc ICD-9 code. As code counts increased, the PPVs increased. PPVs were higher for algorithms using ICD-10-CM codes versus the ICD-9 code. Adding a positive ANA and RP keyword increased the PPVs of algorithms only using ICD billing codes. Algorithms using ??3 or ??4 counts of the SSc ICD-9 or ICD-10-CM codes and ANA positivity had the highest PPV at 100% but a low sensitivity at 50%. The algorithm with the highest F-score of 91% was ??4 counts of the ICD-9 or ICD-10-CM codes with an internally validated PPV of 90%. A machine learning method using random forests yielded an algorithm with a PPV of 84%, sensitivity of 92%, and F-score of 88%. The most important feature was RP keyword. CONCLUSIONS:Algorithms using only ICD-9 codes did not perform well to identify SSc patients. The highest performing algorithms incorporated clinical data with billing codes. EHR-based algorithms can identify SSc patients across a healthcare system, enabling researchers to examine important outcomes.

Project description:Electronic health records (EHRs) have given rise to large and complex databases of medical information that have the potential to become powerful tools for clinical research. However, differences in coding systems and the detail and accuracy of the information within EHRs can vary across institutions. This makes it challenging to identify subpopulations of patients and limits the widespread use of multi-institutional databases. In this study, we leveraged machine learning to identify patterns in medication usage among hospitalized pediatric patients receiving renal replacement therapy and created a predictive model that successfully differentiated between intermittent (iHD) and continuous renal replacement therapy (CRRT) hemodialysis patients. We trained six machine learning algorithms (logistical regression, Naïve Bayes, k-nearest neighbor, support vector machine, random forest, and gradient boosted trees) using patient records from a multi-center database (n = 533) and prescribed medication ingredients (n = 228) as features to discriminate between the two hemodialysis types. Predictive skill was assessed using a 5-fold cross-validation, and the algorithms showed a range of performance from 0.7 balanced accuracy (logistical regression) to 0.86 (random forest). The two best performing models were further tested using an independent single-center dataset and achieved 84-87% balanced accuracy. This model overcomes issues inherent within large databases and will allow us to utilize and combine historical records, significantly increasing population size and diversity within both iHD and CRRT populations for future clinical studies. Our work demonstrates the utility of using medications alone to accurately differentiate subpopulations of patients in large datasets, allowing codes to be transferred between different coding systems. This framework has the potential to be used to distinguish other subpopulations of patients where discriminatory ICD codes are not available, permitting more detailed insights and new lines of research.

Project description:ObjectiveElectronic health records (EHR) offer medical and pharmacogenomics research unprecedented opportunities to identify and classify patients at risk. EHRs are collections of highly inter-dependent records that include biological, anatomical, physiological, and behavioral observations. They comprise a patient's clinical phenome, where each patient has thousands of date-stamped records distributed across many relational tables. Development of EHR computer-based phenotyping algorithms require time and medical insight from clinical experts, who most often can only review a small patient subset representative of the total EHR records, to identify phenotype features. In this research we evaluate whether relational machine learning (ML) using inductive logic programming (ILP) can contribute to addressing these issues as a viable approach for EHR-based phenotyping.MethodsTwo relational learning ILP approaches and three well-known WEKA (Waikato Environment for Knowledge Analysis) implementations of non-relational approaches (PART, J48, and JRIP) were used to develop models for nine phenotypes. International Classification of Diseases, Ninth Revision (ICD-9) coded EHR data were used to select training cohorts for the development of each phenotypic model. Accuracy, precision, recall, F-Measure, and Area Under the Receiver Operating Characteristic (AUROC) curve statistics were measured for each phenotypic model based on independent manually verified test cohorts. A two-sided binomial distribution test (sign test) compared the five ML approaches across phenotypes for statistical significance.ResultsWe developed an approach to automatically label training examples using ICD-9 diagnosis codes for the ML approaches being evaluated. Nine phenotypic models for each ML approach were evaluated, resulting in better overall model performance in AUROC using ILP when compared to PART (p=0.039), J48 (p=0.003) and JRIP (p=0.003).DiscussionILP has the potential to improve phenotyping by independently delivering clinically expert interpretable rules for phenotype definitions, or intuitive phenotypes to assist experts.ConclusionRelational learning using ILP offers a viable approach to EHR-driven phenotyping.

Project description:ImportanceComparisons of antimicrobial use among hospitals are difficult to interpret owing to variations in patient case mix. Risk-adjustment strategies incorporating larger numbers of variables haves been proposed as a method to improve comparisons for antimicrobial stewardship assessments.ObjectiveTo evaluate whether variables of varying complexity and feasibility of measurement, derived retrospectively from the electronic health records, accurately identify inpatient antimicrobial use.Design, setting, and participantsRetrospective cohort study, using a 2-stage random forests machine learning modeling analysis of electronic health record data. Data were split into training and testing sets to measure model performance using area under the curve and absolute error. All adult and pediatric inpatient encounters from October 1, 2015, to September 30, 2017, at 2 community hospitals and 1 academic medical center in the Duke University Health System were analyzed. A total of 204 candidate variables were categorized into 4 tiers based on feasibility of measurement from the electronic health records.Main outcomes and measuresAntimicrobial exposure was measured at the encounter level in 2 ways: binary (ever or never) and number of days of therapy. Analyses were stratified by age (pediatric or adult), unit type, and antibiotic group.ResultsThe data set included 170 294 encounters and 204 candidate variables from 3 hospitals during the 3-year study period. Antimicrobial exposure occurred in 80 190 encounters (47%); 64 998 (38%) received 1 to 6 days of therapy, and 15 192 (9%) received 7 or more days of therapy. Two-stage models identified antimicrobial use with high fidelity (mean area under the curve, 0.85; mean absolute error, 1.0 days of therapy). Addition of more complex variables increased accuracy, with largest improvements occurring with inclusion of diagnosis information. Accuracy varied based on location and antibiotic group. Models underestimated the number of days of therapy of encounters with long lengths of stay.Conclusions and relevanceModels using variables derived from electronic health records identified antimicrobial exposure accurately. Future risk-adjustment strategies incorporating encounter-level information may make comparisons of antimicrobial use more meaningful for hospital antimicrobial stewardship assessments.