Background
Hypertrophic cardiomyopathy (HCM) is the most common inherited heart disease and a leading cause of sudden cardiac death (SCD) in young adults. Despite existing clinical practice guidelines, HCM cases remain underrecognized and sub-optimally managed. Studies using large electronic health record (EHR) EHR-based cohorts of HCM patients are necessary to fulfill existing HCM knowledge gaps. Cohort identification using traditional methods, including manual chart abstraction of EHR data, reading cardiac images and genetic reports, and manual review of clinical notes by an expert or well-trained clinician, are time-consuming and very expensive. To overcome these barriers, we propose a new approach for accurate and efficient cohort identification using an artificial intelligence (AI), machine learning (ML) model, which can automatically identify HCM patients using echocardiogram (echo) reports as a data source.
Methods
The current study used structured data from two HCM cohorts, one composed of 5,835 adult patients and the second with 277 children, with at least one diagnostic billing code for HCM between January 2004 and March 2019. The institutional review board approved this study. Patients who declined research authorization were excluded. Billing codes used were 425.11 and 425.18 (ICD-9) and I42.1, and I42.2 (ICD-10). All patients had confirmed HCM diagnosis by reports of clinically indicated echocardiograms. Three control patients without billing codes for HCM were matched by age, gender, and the year of echo report with one HCM patient. Echo report data were subsequently retrieved from the institutional data warehouse. In the next step, clinicians and data scientists focused on selecting the relevant diagnostic features to extract HCM-related echo measurements and impressions from a list of over 900 measurements and over 4000 impressions. The selected measurements and impressions were combined and narrowed down into 60 HCM-related measurements and ten categorical impressions. Random forest was then used to investigate the predictive performance of the selected features for identifying HCM patients due to practical application and expected accuracy in a wide range of use cases. This model aimed to reduce the complexity with a simple voting strategy while preventing the overfitting problem. The selected echo measurements and the refined impressions were used as input for the ML models to train two models for children and two models for adults: echo measurement-based models and echo measurement and impression-based models. In the next step, the data set was divided into training and test dataset to train each model to predict a label for any new test sample using majority votes. The outcomes for these models were HCM diagnosis labels YES or NO. 10-fold cross-validation was applied with 80% random selection for the train and 20% for the test data set at each round to validate the results. The most important measurements and impressions for each model were extracted and reviewed by cardiologists and sonographers to assure model interpretability .
Results
Our ML models had a high accuracy of 94% for measurement-based models and 99% for measurement and impression-based models for pediatric and adult cohorts. The positive predictive value of measurement-based models was 93%, and for the measurement and impression-based models, it was 99%. The top 20 features of each model matched well with the clinician's thought process for echo interpretation. In all models, diastolic dysfunction parameters, left ventricular outflow tract obstruction, interventricular septum diastolic thickness, left ventricular mass index, ejection fraction, and left ventricular longitudinal strain were among the top ten predictors. Differences between models for children and adults included the importance order of the top ten predictors.
Conclusion
ML models developed from echo measurements and impression identify HCM cases with high performance. These models will be used for clinical research, quality improvement projects and to find the right patients to populate clinical decision support systems.

**Background**
Hypertrophic cardiomyopathy (HCM) is the most common inherited heart disease and a leading cause of sudden cardiac death (SCD) in young adults. Despite existing clinical practice guidelines, HCM cases remain underrecognized and sub-optimally managed. Studies using large electronic health record (EHR) EHR-based cohorts of HCM patients are necessary to fulfill existing HCM knowledge gaps. Cohort identification using traditional methods, including manual chart abstraction of EHR data, reading cardiac images and genetic reports, and manual review of clinical notes by an expert or well-trained clinician, are time-consuming and very expensive. To overcome these barriers, we propose a new approach for accurate and efficient cohort identification using an artificial intelligence (AI), machine learning (ML) model, which can automatically identify HCM patients using echocardiogram (echo) reports as a data source.
**Methods**
The current study used structured data from two HCM cohorts, one composed of 5,835 adult patients and the second with 277 children, with at least one diagnostic billing code for HCM between January 2004 and March 2019. The institutional review board approved this study. Patients who declined research authorization were excluded. Billing codes used were 425.11 and 425.18 (ICD-9) and I42.1, and I42.2 (ICD-10). All patients had confirmed HCM diagnosis by reports of clinically indicated echocardiograms. Three control patients without billing codes for HCM were matched by age, gender, and the year of echo report with one HCM patient. Echo report data were subsequently retrieved from the institutional data warehouse. In the next step, clinicians and data scientists focused on selecting the relevant diagnostic features to extract HCM-related echo measurements and impressions from a list of over 900 measurements and over 4000 impressions. The selected measurements and impressions were combined and narrowed down into 60 HCM-related measurements and ten categorical impressions. Random forest was then used to investigate the predictive performance of the selected features for identifying HCM patients due to practical application and expected accuracy in a wide range of use cases. This model aimed to reduce the complexity with a simple voting strategy while preventing the overfitting problem. The selected echo measurements and the refined impressions were used as input for the ML models to train two models for children and two models for adults: echo measurement-based models and echo measurement and impression-based models. In the next step, the data set was divided into training and test dataset to train each model to predict a label for any new test sample using majority votes. The outcomes for these models were HCM diagnosis labels YES or NO. 10-fold cross-validation was applied with 80% random selection for the train and 20% for the test data set at each round to validate the results. The most important measurements and impressions for each model were extracted and reviewed by cardiologists and sonographers to assure model interpretability .
**Results**
Our ML models had a high accuracy of 94% for measurement-based models and 99% for measurement and impression-based models for pediatric and adult cohorts. The positive predictive value of measurement-based models was 93%, and for the measurement and impression-based models, it was 99%. The top 20 features of each model matched well with the clinician's thought process for echo interpretation. In all models, diastolic dysfunction parameters, left ventricular outflow tract obstruction, interventricular septum diastolic thickness, left ventricular mass index, ejection fraction, and left ventricular longitudinal strain were among the top ten predictors. Differences between models for children and adults included the importance order of the top ten predictors.
**Conclusion**
ML models developed from echo measurements and impression identify HCM cases with high performance. These models will be used for clinical research, quality improvement projects and to find the right patients to populate clinical decision support systems.

Automated Machine Learning Models Based on Echocardiographic Features can Improve HCM Cohort Identification for Artificial Intelligence Applications