Background
Hypertrophic Cardiomyopathy (HCM) is the most common heritable cardiomyopathy in the United States, with an estimated prevalence of 1 in 500 and the most frequent cause of sudden cardiac death in young adults. The gold standard for HCM diagnosis is echocardiography (echo), an ultrasound-based cardiac imaging modality. To establish the diagnosis of HCM, an enormous amount of information from echo images is acquired and analyzed. In the current study, we have studied the common causal factors in the delayed or missed opportunities for HCM diagnosis by echo. We compared historical billing codes against machine learning (ML) techniques to identify areas for improvement. We hypothesized that automated methodologies for data processing using ML technology in real-time would greatly increase clinical workflow efficiency while ensuring high-quality interpretation, which is a fundamental requirement for accurate HCM diagnosis.
Methods
The Mayo Clinic Cardiovascular Ultrasound Imaging and Hemodynamic Laboratory (Echo Lab) is the largest clinical, educational, and research echocardiographic laboratory facility of its kind in the world. We retrospectively studied a large echo dataset covering 10,818 patients, including 5,610 patients with confirmed echocardiographic diagnosis of HCM (definite HCM). The date of the first echo report with impressions indicating HCM diagnosis was considered the index date. Data from each patient's first available echo report was selected to evaluate if machine learning algorithms could enable earlier identification of HCM patients. These algorithms were specifically designed to assist providers in an efficient and accurate diagnosis of HCM by automated phenotyping of echo lab data. A total of 150 measurements and impressions (coded sentences describing categorical variables from the reports) were extracted from the first echo report of each patient and passed through various preprocessing steps to handle two main barriers a) replicated and deprecated measurements and (b) frequent missing measurements. The problem of replicated and deprecated measurements appears specifically in longitudinal studies where the introduction of new technologies resulted in adding measurements and impressions (e.g., 3D mode) to existing data (e.g., 2D mode). Such transitions in practice cause some variables to be deprecated and discontinued, while new measurements beholding part of a similar concept. Our partial solution for this problem was simplified by combining complementary variables into a single comprehensive concept. This way, we moved from a total of 150 variables to 60 variables of less sparsity. After data cleaning, we trained a random forest on a random 80% subsample of data and tested it against the remaining data samples. The recent manual clinical chart reviews were used as the ground truth in evaluating the performance of ML models for the identification of HCM patients. We report a missed or delayed opportunity for diagnosis if the ML model could correctly identify the patient for HCM (in their first echo visit) when there is no confirmatory HCM diagnosis in the patient echo report in the EHR by that date. We can also compute the time delay between the time this ML model identifies HCM patients (date of their first echo) and the date that HCM diagnosis was reported for the first time in her echo report.
Results
The ML models showed high performance (accuracy 94%, sensitivity 98% for measurement-based models and accuracy 99%, sensitivity 99% for measurement, and impression-based models for identification of HCM patients. Additionally, this model has also identified that 20% of the HCM definite cases earlier than the HCM diagnosis were reported by echo (average 41 ± 45 months). A subset of 76 patients had confirmed HCM diagnosis in the echo reports ten years after their first echoes. Additionally, almost half of important HCM measurements were not filled in on the date of the first echo reports. A potential contributor to this was the adoption of new echo technologies with consequently need to incorporate new information in echo reports. The absence of mention of HCM diagnosis in the first echoes indicates a need for standardization of the content of echo reports. Machine learning models can be used during imaging acquisition, interpretation, and reporting phases of the echo laboratory workflow to standardize clinical processes and assist providers in diagnosing HCM. Additionally, this technology may be used in Echo labs that serve rural and underserved communities with less experience diagnosing HCM, and for learners (student sonographers, cardiology fellows) who may also benefit from potential clinical application for this technology. Primary care providers who use the results of echo reports for patient management may also benefit from the adoption of standardization of echo reports assisted by technology with a clear understanding of the patient diagnosis available in the echo report.
Conclusion
The echo reports of over five thousand patients with definite HCM were studied and revealed that in 20% of these reports, the HCM diagnosis was not clearly specified. The recent ML models can find HCM from the first echo of an HCM patient with high accuracy. Therefore, the standardization of echo reports based on HCM-specific measurements can enable prompt reports of HCM diagnosis. Sonographers and clinicians would get real-time feedback about HCM diagnosis after these models are deployed for use by providers in echo laboratories. We anticipate that by using Machine Learning-Based Decision Support Systems, sonographers and physicians would be reminded to include measurements necessary to establish HCM diagnosis in echo reports.

**Background**
Hypertrophic Cardiomyopathy (HCM) is the most common heritable cardiomyopathy in the United States, with an estimated prevalence of 1 in 500 and the most frequent cause of sudden cardiac death in young adults. The gold standard for HCM diagnosis is echocardiography (echo), an ultrasound-based cardiac imaging modality. To establish the diagnosis of HCM, an enormous amount of information from echo images is acquired and analyzed. In the current study, we have studied the common causal factors in the delayed or missed opportunities for HCM diagnosis by echo. We compared historical billing codes against machine learning (ML) techniques to identify areas for improvement. We hypothesized that automated methodologies for data processing using ML technology in real-time would greatly increase clinical workflow efficiency while ensuring high-quality interpretation, which is a fundamental requirement for accurate HCM diagnosis.
**Methods**
The Mayo Clinic Cardiovascular Ultrasound Imaging and Hemodynamic Laboratory (Echo Lab) is the largest clinical, educational, and research echocardiographic laboratory facility of its kind in the world. We retrospectively studied a large echo dataset covering 10,818 patients, including 5,610 patients with confirmed echocardiographic diagnosis of HCM (definite HCM). The date of the first echo report with impressions indicating HCM diagnosis was considered the index date. Data from each patient's first available echo report was selected to evaluate if machine learning algorithms could enable earlier identification of HCM patients. These algorithms were specifically designed to assist providers in an efficient and accurate diagnosis of HCM by automated phenotyping of echo lab data. A total of 150 measurements and impressions (coded sentences describing categorical variables from the reports) were extracted from the first echo report of each patient and passed through various preprocessing steps to handle two main barriers a) replicated and deprecated measurements and (b) frequent missing measurements. The problem of replicated and deprecated measurements appears specifically in longitudinal studies where the introduction of new technologies resulted in adding measurements and impressions (e.g., 3D mode) to existing data (e.g., 2D mode). Such transitions in practice cause some variables to be deprecated and discontinued, while new measurements beholding part of a similar concept. Our partial solution for this problem was simplified by combining complementary variables into a single comprehensive concept. This way, we moved from a total of 150 variables to 60 variables of less sparsity. After data cleaning, we trained a random forest on a random 80% subsample of data and tested it against the remaining data samples. The recent manual clinical chart reviews were used as the ground truth in evaluating the performance of ML models for the identification of HCM patients. We report a missed or delayed opportunity for diagnosis if the ML model could correctly identify the patient for HCM (in their first echo visit) when there is no confirmatory HCM diagnosis in the patient echo report in the EHR by that date. We can also compute the time delay between the time this ML model identifies HCM patients (date of their first echo) and the date that HCM diagnosis was reported for the first time in her echo report.
**Results**
The ML models showed high performance (accuracy 94%, sensitivity 98% for measurement-based models and accuracy 99%, sensitivity 99% for measurement, and impression-based models for identification of HCM patients. Additionally, this model has also identified that 20% of the HCM definite cases earlier than the HCM diagnosis were reported by echo (average 41 ± 45 months). A subset of 76 patients had confirmed HCM diagnosis in the echo reports ten years after their first echoes. Additionally, almost half of important HCM measurements were not filled in on the date of the first echo reports. A potential contributor to this was the adoption of new echo technologies with consequently need to incorporate new information in echo reports. The absence of mention of HCM diagnosis in the first echoes indicates a need for standardization of the content of echo reports. Machine learning models can be used during imaging acquisition, interpretation, and reporting phases of the echo laboratory workflow to standardize clinical processes and assist providers in diagnosing HCM. Additionally, this technology may be used in Echo labs that serve rural and underserved communities with less experience diagnosing HCM, and for learners (student sonographers, cardiology fellows) who may also benefit from potential clinical application for this technology. Primary care providers who use the results of echo reports for patient management may also benefit from the adoption of standardization of echo reports assisted by technology with a clear understanding of the patient diagnosis available in the echo report.
**Conclusion**
The echo reports of over five thousand patients with definite HCM were studied and revealed that in 20% of these reports, the HCM diagnosis was not clearly specified. The recent ML models can find HCM from the first echo of an HCM patient with high accuracy. Therefore, the standardization of echo reports based on HCM-specific measurements can enable prompt reports of HCM diagnosis. Sonographers and clinicians would get real-time feedback about HCM diagnosis after these models are deployed for use by providers in echo laboratories. We anticipate that by using Machine Learning-Based Decision Support Systems, sonographers and physicians would be reminded to include measurements necessary to establish HCM diagnosis in echo reports.

Machine Learning-Based Decision Support Systems for Standardization of Echocardiogram Reports