Characterization of rat vertebrae cortical bone microstructures using confocal Raman microscopy combined to tomography and electron microscopy

BACKGROUND

The rat vertebrae is a good model to study bone regeneration after implantation of biomaterials used to treat bone loss, a major problem in oral and dental surgery. However, the precise characterization of bone microstructures in the rat vertebrae has not been reported. Therefore, the aim of this study was to achieve the complete analysis of such bone, at different scales, in order to have a clear model of healthy bone for comparison with regenerated bone.

METHODS

In order to image the cortical bone of rat caudal vertebra, confocal Raman microscopy was combined with high resolution X-ray micro computed tomography (micro-CT), with scanning electron microscopy (SEM) using backscatter electron imaging and with more conventional histology coloration techniques. SEM and Raman microscopy were done in various regions of the cortical bone corresponding to external, middle and internal areas. The spongy bone was imaged in parallel. Micro-CT was performed on the whole vertebra to monitor the network of haversian canals in the cortical bone. Osteonic canals characteristics, and relative chemical composition were analysed in several regions of interest, in cortical and spongy bone. Five rats were included in this study.

RESULTS

On micro-CT images, differences in intensity were observed in the cortical bone, substantiated by SEM. Chemical analysis with Raman spectra confirmed the difference in composition between the different regions of the cortical and spongy bone. PCA and k-mean cluster analysis separated these groups, except for the external and middle cortical bone. Peak intensity ratio confirmed these results with a CO3 to ν2 PO4 ratio significantly different for the internal cortical bone. Grayscale images stack extracted from micro-CT showed that global architecture of cortical bone was characterized by a dense and complex network of haversian osteonic canals, starting from the surface towards the vertebrae center. The mean diameter of the canals was 18.4 µm (SD 8.6 µm) and the mean length was 450 µm (SD 152 µm). Finally, Raman reconstructed images of the lamellar bone showed an enlargement of the lamellar layer width, both in circumferential lamellar bone and around haversian canals.

CONCLUSIONS

Micro-CT and confocal Raman microscopy are good tools to complete classical analysis using optical and electron microscopy. The results and measurements presented in a rat model known for its small inter-individual differences provide the main characteristics of a mature bone. This study will allow the community working on this rat vertebrate model to have a set of characteristics, in particular on the structure of the haversian canals.

Engaged with the heart – the EKG ring STEMI detector

Background

Our previous experience in telemedicine-centered STEMI management networks has shown delayed presentations as one of the most relevant factors in the reduction of Symptom-to-Baloon times. In order to further improve outcomes, we delved into the applications of mathematical vector algebra and engineering to incorporate an innovative Artificial Intelligence-guided Single Lead EKG methodology into a wearable ring to provide a self-administered alternative to reliable and expedite STEMI screening.

Purpose

To provide preliminary results of the application of ultra-wearable technology in a ring for accurate STEMI detection.

Methods

Our present work was done in two steps – 1) Applying mathematical vector algebra to construct an accurate and practical AI-guided Single Lead EKG algorithm for STEMI detection compatible with wearable devices, and 2) To engineer this algorithm into a wearable ring for quick and reliable STEMI detection. Throughout our first step, we provided a group of new lead waveforms (Vn') by positioning a single lead-capable wearable device into the chest positions Cn (C1, C2,..., C6) while touching the second electrode with a right-hand finger in the same device, all of which corresponded to the difference in electric potential between Right Arm (RA) and the correspondent conventional precordial Vn chest position. By using vector algebra, we recognized Vn' as the sum of (-aVR + Vn). Vector mathematical analysis was performed for 5,783 STEMI (50%) and 5,784 Not-STEMI (50%) EKG from a proprietary dataset, obtaining their corresponding new Vn' precordial leads. Finally, the AI-guided STEMI detector model was trained with 10,410 EKG records (90%) and tested with 1,157 EKG records (10%). Performance metrics were calculated to determine best new Lead for STEMI detection. In the second step, we engineered this methodology into a wearable ring device. When a patient presents chest discomfort or oppression, the most common reaction is to move the hands towards the chest. By mimicking this behavior and having our EKG-capable ring technology on the right hand, we replicate our methodology by positioning said ring to chest positions Cn to register an EKG trace of new Vn' precordial leads and calculated performance metrics to evaluate the correlation with previous experiment.

Results

Test results shows Lead V2' as the best overall lead in detecting STEMI with 91.2% Accuracy, 89.6% Sensibility, and 92.9% Specificity. These results were reproduced with both methodologies.

Conclusions

Preliminary outcomes of the implementation of our innovative Single Lead EKG methodology into an ultra-portable ring yielded promising results. Prospective studies will be needed to further validate this neoteric methodology for STEMI detection, nevertheless, we envision the potential future applications of this technology in the clinical setting, particularly with swift screening and activation of remote STEMI management networks.

Funding Acknowledgement

Type of funding source: None

Reconfiguring traditional EKG interpretation with artificial intelligence – a reliable, time-saving alternative?

Background

Time and accuracy are key factors that may make or break an efficient triage and management in most medical premises, particularly so when expedited diagnosis saves lives - a not so uncommon scenario in the field of cardiology. By studying the different variables involved in cardiologist-EKG interactions that lead to the identification and management of different cardiovascular entities, we delved into the applications of Artificial Intelligence (AI) algorithms in order to improve upon the classic, but dated, EKG methodology. With this study, we pit our algorithm against cardiologists to perform a thorough analysis of the time invested to diagnose an EKG as Normal, as well as an assessment of the accuracy of said label.

Purpose

To present a faster and reliable AI-guided EKG interpretation methodology that outperforms cardiologists' capabilities in identifying Normal EKG records.

Methods

The International Telemedical System (ITMS) developed and tested an EKG assessing AI algorithm and incorporated it into the workflow of their Telemedicine Integrated Platform, a digital EKG reading program where cardiologists continuously report their findings remotely in real time. During the month of April 2019; 35 ITMS cardiologists reported a grand total of 61,441 EKG records, later subjecting them to the AI algorithm, implemented through the “One Click Report” process. Through this simple 2-step approach, the algorithm provides a suggestion of “Normal” or “Abnormal” to the cardiologist based on the patterns of the fiducial points included in said EKG reports. A comparison of the time of normal EKG diagnosis is made and the correlation between AI and cardiologists is assessed.

Results

On average, our AI algorithm discerned a normal EKG within 30.63s (95% CI 26.51s to 34.75s), in solid contrast with cardiologists' interpretations alone, which amounted to 83.54s (95% CI from 69.43s to 97.65s). This accounts for an overall saving of 52.91s (95% CI 42.45s to 63.83s) by implementing this innovative methodology in a cardiologist practice. In addition, this method correctly reported 23,213 Normal EKG records out of the total 25,013 AI output, reaching a 92.8% correlation between man and machine. The total average time saved in normal EKG readings with AI (23,213) would accrue an approximate of 20,470 minutes (ie, 42 8-hours work shifts worth of time dedicated to diagnosing a normal EKG).

Conclusions

The implementation of automated AI-driven technologies into daily EKG interpretation tasks poses an attractive time-saving alternative for faster and accurate results in a modern cardiology practice. By further expanding on the concept of an intelligent EKG characterization device, a more efficient and patient-centered clinical exercise will ensue.

Funding Acknowledgement

Type of funding source: None

Enhancing AI-guided STEMI detection algorithms by incorporating higher quality fiduciary EKG elements

Background

As EKG interpretation paradigms to a physician-free milieu, accumulating massive quantities of distilled pre-processed data becomes a must for machine learning techniques. In our pursuit of reducing ischemic times in STEMI management, we have improved our Artificial Intelligence (AI)-guided diagnostic tool by following a three-step approach: 1) Increase accuracy by adding larger clusters of data. 2) Increase the breadth of EKG classifications to provide more precise feedback and further refine the inputs which ultimately reflects in better and more accurate outputs. 3) Improving the algorithms' ability to discern between cardiovascular entities reflected in the EKG records.

Purpose

To bolster our algorithm's accuracy and reliability for electrocardiographic STEMI recognition.

Methods

Dataset: A total of 7,286 12-lead EKG records of 10-seconds length with a sampling frequency of 500 Hz obtained from Latin America Telemedicine Infarct Network from April 2014 to December 2019. This included the following balanced classes: angiographically confirmed STEMI, branch blocks, non-specific ST-T abnormalities, normal, and abnormal (200+ CPT codes, excluding the ones included in other classes). Labels of each record were manually checked by cardiologists to ensure precision (Ground truth). Pre-processing: First and last 250 samples were discarded to avoid a standardization pulse. Order 5 digital low pass filters with a 35 Hz cut-off was applied. For each record, the mean was subtracted to each individual lead. Classification: Determined classes were “STEMI” and “Not-STEMI” (A combination of randomly sampled normal, branch blocks, non-specific ST-T abnormalities and abnormal records – 25% of each subclass). Training & Testing: A 1-D Convolutional Neural Network was trained and tested with a dataset proportion of 90/10, respectively. The last dense layer outputs a probability for each record of being STEMI or Not-STEMI. Additional testing was performed with a subset of the original complete dataset of unconfirmed STEMI. Performance indicators (accuracy, sensitivity, and specificity) were calculated for each model and results were compared with our previous findings from past experiments.

Results

Complete STEMI data: Accuracy: 95.9% Sensitivity: 95.7% Specificity: 96.5%; Confirmed STEMI: Accuracy: 98.1% Sensitivity: 98.1% Specificity: 98.1%; Prior Data obtained in our previous experiments are shown below for comparison.

Conclusion(s)

After the addition of clustered pre-processed data, all performance indicators for STEMI detection increased considerably between both Confirmed STEMI datasets. On the other hand, the Complete STEMI dataset kept a strong and steady set of performance metrics when compared with past results. These findings not only validate the consistency and reliability of our algorithm but also connotes the importance of creating a pristine dataset for this and any other AI-derived medical tools.

Funding Acknowledgement

Type of funding source: None

Waddling beyond door to balloon times and impinging true ischemic times with artificial intelligence-guided single lead EKG for STEMI detection

Background

The present process of STEMI detection is cumbersome as it utilizes outdated equipment and requires a trained technician and an expert cardiologist. We have developed a patient-administered, Artificial Intelligence (AI) guided, Single Lead EKG for early STEMI detection.

Purpose

To answer the question “Is early STEMI detection possible with a Single Lead EKG?”

Methods

We experimented with an AI-guided algorithm for a single-lead EKG for STEMI detection with the following step-wise developments: 1) An AI algorithm that predictably interprets STEMI using a 12-lead EKG; 2) An AI algorithm for STEMI detection using a single-lead EKG; 3) A methodology for identifying the best single lead to detect STEMI; 4) Advanced AI algorithms for STEMI localization with a single-lead EKG. The AI methodology was as follows: Sample: The mammoth Latin American Telemedicine Infarct Network telemedicine database that provides an umbrella of AMI management to 100 million patients in Brazil, Colombia, Mexico, Chile, and Argentina was queried for cardiologist annotated EKG. A total of 8,511 EKG and 90,592 classified heartbeats were selected for the experiments. Preprocessing: segmentation of each ECG into individual heartbeats. Training & Testing: 90% and 10%, respectively, of the total dataset. Classification: 1-D Convolutional Neural Network; classes were construed for each heartbeat. Performance indicators were calculated per lead.

Results

The algorithm was able to provide an accuracy of 91.9%. Lead V2 yielded the best results among individual leads for STEMI detection.

Conclusions

Early experiments provide a framework for augmenting STEMI detection with the use of AI-guided, single lead techniques. Such approaches seem rational as we target the reduction of true STEMI ischemic times.

Funding Acknowledgement

Type of funding source: None

Countdown to physician-free EKG interpretation

Background

With the introduction of electronic medical records and other digital platforms, the classification and coding of different medical entities have become a complex, cumbersome task that is prone to diagnostic inconsistencies and errors. By incorporating Artificial Intelligence (AI) to a massive database of EKG records, we have developed an innovative methodology to accurately discriminate an EKG as “normal” or “abnormal”. We firmly believe that this algorithm sets up medicine on a path of complete computer-aided EKG interpretation.

Purpose

To present a viable AI-guided filter that can accurately discriminate between normal and abnormal EKG within a cardiologist-annotated EKG database.

Methods

An observational, retrospective, case-control study. Samples: A total of 140,000 randomly sampled 12-lead ECG of 10-seconds length with a sampling frequency of 500 [Hz] from Brazil (BR) and Colombia (CO) (divided as 70,000 normal and 70,000 abnormal EKG records per country dataset) were derived from the private International Telemedical System (ITMS) database from September 2018 to July 2019. Only de-identified records were used, records with artifacts were excluded. Preprocessing: Only the first 2s of each short lead and 9s of the long lead were considered. This data includes mobile (MOB) and transtelephonic (TTP) EKGs (50/50 ratio). Limb leads I, II and III and precordial leads V1, V2, V3 and V5 were used. The mean was removed from each lead. Training Sets: Four models were trained as depicted in the figure below. Each training dataset has 25,000 Normal and 25,000 Abnormal records, where 10% of the total records were used as a validation set. The test sets included 10,000 normal, and 10,000 abnormal records each. Testing and Class Assigning: An inception convolutional neural network was implemented; Each model was tested with 5,000 normal and 5,000 abnormal records of the corresponding country and transmission type with which they were trained. “Normal” or “Abnormal” labels were assigned to each EKG record and were compared to the cardiologists' reports; performance indicators (accuracy, sensitivity, and specificity) were calculated for each model.

Results

An overall accuracy of 82.4%; sensitivity of 88.7%; and specificity of 76.2% was achieved amongst the 4 testing models (Separate results of each training set are shown below).

Conclusion(s)

AI enables the interpretation of digital EKG records to be exercised in an organized, accurate, and straightforward manner, taking into consideration the multiple potential entities that can be diagnosed through this historical triage tool. By quickly identifying the normal records, the cardiologist is able to invest efforts in treating patients in a timely manner.

Funding Acknowledgement

Type of funding source: None

Maximum artificial intelligence and complete reconstruct of population-based AMI care

Background

After creating a behemoth hub and spoke AMI network that encompasses more than 100 million patients in 5 countries, we have begun to incorporate Artificial Intelligence (AI) algorithms into our telemedicine strategy with the goal of creating comprehensive, very early AMI diagnosis and physician-free triage. In doing so, we have replaced door-to-balloon times (d2b) with symptom-to-balloon times (s2b) as an immutable objective.

Purpose

To incorporate AI attributes for very early AMI detection, triage, and management.

Methods

We expanded our effective telemedicine strategy (100 million population; 877,178 telemedicine encounters; 55% overall mortality reduction; $291 million cost savings) with a logistic reset to impact s2b. To do this, we incorporated our Single Lead 1.0 (lead I) and Single Lead 2.0 (lead V2) technology for self-administered AMI detection with our physician-free STEMI diagnosis and triage AI algorithms. Single Lead algorithms and physician-free protocols were generated by utilizing Machine Learning from our mammoth annotated EKG repository.

Results

In addition to three logistic markers of efficiency Time-to-Telemedicine Diagnosis (TTD), Door-In-Door-Out (DIDO) and Transfer Times (TT); we are monitoring s2b. A gradual release of the algorithms and single lead is occurring at the telemedicine spokes. Detailed results will be available at the time of presentation.

Conclusions

Impacting s2b, the Achilles Heel of Primary PCI, may be achieved with the use of patient-administered AMI detection tools. Incorporation of these technologies into AI algorithms will add to telemedicine efficiencies for population-based AMI care.

Funding Acknowledgement

Type of funding source: None

AI and telemedicine: total remote guidance of AMI management

Background

For a decade, Latin American Telemedicine Infarct Network (LATIN) Telemedicine has transformed AMI management in Brazil, Colombia, Mexico, Chile, and Argentina. With a hub and spoke strategy, AMI coverage was expanded to 100 million population and 877,177 telemedicine encounters were performed. Cost savings from avoiding unnecessary transfer of patients was $291 million. We are now rapidly escalating on a path to making the telemedicine process “physician-free” by utilizing Artificial Intelligence (AI) protocols.

Purpose

To demonstrate that AI can replace a cardiologist for remote AMI telemedicine guidance.

Methods

The process of total AI guidance focused on both aspects of our telemedicine strategy – accurate AMI diagnosis and tele-guidance of the entire STEMI process. We developed our innovative approach by initially creating AI algorithms for computer-aided diagnosis. Next, we incorporated logistic variables (duration of chest pain, transfer times to LATIN hub, etc) to the algorithm for physician-free triage into thrombolysis, primary PCI and pharmaco-invasive management. The intent of creating AI algorithms was early STEMI detection and triage. After the patient was efficiently transferred to the hub, a final treatment decision was made by the hub cardiologists.

Results

Three crucial areas of telemedicine efficiency are being monitored – Time-to-Telemedicine Diagnosis (TTD), Door-In-Door-Out (DIDO) and Transfer Times (TT). All are showing improvements. Detailed results will be available at the time of presentation.

Conclusions

We are encouraged with the possibility of making the entire telemedicine guidance of AMI management “physician-free”. Next-Gen improvements are being contemplated by including a Single Lead EKG for AMI detection that will impact symptom-to-balloon times.

Funding Acknowledgement

Type of funding source: None

Moving in sync – concordance betweena artificial intelligence and cardiologist on detecting normal electrocardiograms

Background

Merging modern technologies with classic diagnostic tests often results in a sense of insecurity within the medical community, particularly so with potentially life-saving studies such as the electrocardiogram (EKG). In order to provide a greater sense of trust between Artificial Intelligence (AI) and cardiologists, we provide an AI-driven algorithm capable of accurately and reliably characterize an EKG as normal within a highly complex, cardiologist-reviewed EKG database and report the degree of concordance between this machine vs physician scenario.

Purpose

To provide a dependable and accurate AI algorithm that conducts EKG interpretation in a cardiologist-tier manner.

Methods

The International Telemedical System (ITMS) developed and tested an EKG assessing AI algorithm and incorporated it into the workflow of their Telemedicine Integrated Platform, a digital EKG reading program where cardiologists continuously report their findings remotely in real-time. During the month of April 2,019; 35 ITMS cardiologists reported a grand total of 61,441 EKG records, later submitting them to the AI algorithm implemented through the “One Click Report” process. Through this simple 2-step approach, the algorithm provides a suggestion of “Normal” or “Abnormal” to the cardiologist based on the patterns of the fiducial points included in said EKG reports. Confirmation of these suggestions by the cardiologists ensued.

Results

Overall, cardiologists confirmed 23,213 out of 25,013 AI outputs for “Normal” EKGs, demonstrating a concordance of 92.8% for Normal diagnosis.

Conclusion

Through this methodology, we provide an AI technology that can be reliably applied and trusted in EKG digital platforms to identify and suitably label a normal EKG. Further testing will accrue into a multi label algorithm compatible with abnormal cardiovascular entities, potentially precluding the role of the cardiologist for triaging, particularly in the prehospital setting. We anticipate that this approach will become a promising methodology in modern cardiology practice.

Funding Acknowledgement

Type of funding source: None

Key changes in indices of myocardial work in cardiometabolic disease states

Background

Hypertension (HTN) and diabetes mellitus (DM) are prevalent cardiovascular disease states which have been shown to be associated with adverse cardiac remodelling and subclinical myocardial dysfunction. Myocardial work (MW) indices are novel non-invasive measures of left ventricular (LV) function. We aimed to characterise key differences in MW indices in patients with these conditions.

Methods

Outpatients with HTN and DM undergoing transthoracic echocardiography (TTE; 2016–2019) at our institution were assessed and compared to healthy controls. Only patients without cardiac disease with normal diastolic parameters on TTE were recruited. Patients with impaired LV function, cardiac ischaemia or arrhythmia, structural and valvular heart disease or poor-quality images were excluded. Recruited patients were stratified into 3 groups (Group1: Healthy Controls; Group 2: HTN; Group 3: HTN-DM). MW assessment was performed using GE E-95 EchoPac v2.2 system.

Results

Three hundred patients (57.3±17.4y, 51% female) were recruited. HTN and HTN-DM patients were associated with higher resting systolic blood pressure (SBP), indexed LV mass (LVMI), e' and E/e' compared to controls but no differences were noted in these parameters between HTN and HTN-DM. Global myocardial work index (GWI) was higher in HTN patients compared to Controls but not different compared to HTN-DM. Of interest, HTN-DM patients had higher global myocardial wasted work (GWW) and lower global myocardial work efficiency (GWE) compared to HTN patients and Controls.

Conclusions

MW indices may be a sensitive tool for the detection of subclinical changes in cardiac function in cardiometabolic disease states.

Funding Acknowledgement

Type of funding source: None

Enriching artificial intelligence ST-elevation myocardial infarction (STEMI) detection algorithms with differential diagnoses

Background

STEMI outcomes, although improved with systems of care, are hamstrung by delayed presentation and prevaricates of a 12-lead ECG. We report an artificial intelligence (AI) guided, single lead EKG algorithm for a self-administered tool to reliably detect STEMI and trigger ambulance dispatch.

Purpose

To provide a reliable and improved AI-guided Single Lead EKG methodology.

Methods

From our cardiologist-annotated repository, we assigned a dataset of 11,118 classified ECG. Ontology organized 5 groups apportioned for an interclass balance among commoner STEMI differential diagnoses. This anonymous, pre-classified data included 5,549 STEMI, 1,391 normal, 1,393 Bundle Branch Block, 1,393 non-specific ST-T changes and 1,392 miscellaneous. Each ECG was fragmented into individual 1-lead strips. Algorithm: 1-D Convolutional Neural Networks. Gender and age were included before the last dense layer. Training and Testing: Preset 90% dataset (10,008 ECG) train, 10% test (1,110 ECG). Statistical Analysis and ROC curves: Digitized dataset, 500 samples/second, 10s duration, total 5,000 samples per lead. Statistical mean for each lead was calculated and subtracted from the original lead. Statistical values and ROC curves were assessed.

Results

Most Accurate: Lead V2 – 91%; Most Sensitive: Lead I – 92% Most Specific: Lead III – 96%. Best AUC: Lead V2 – 91%.

Conclusions

Incorporating subtypes of STEMI differential diagnosis enriches the single lead AI algorithm. Validating the derived algorithm with our entire database of 18 million ECG will further strengthen the results.

Funding Acknowledgement

Type of funding source: None

Baby steps in the path of modifying the role of cardiologists for interpreting EKG for AMI

Background

EKG interpretation is slowly transitioning to a physician-free, Artificial Intelligence (AI)-driven endeavor. Our continued efforts to innovate follow a carefully laid stepwise approach, as follows: 1) Create an AI algorithm that accurately identifies STEMI against non-STEMI using a 12-lead EKG; 2) Challenging said algorithm by including different EKG diagnosis to the previous experiment, and now 3) To further validate the accuracy and reliability of our algorithm while also improving performance in a prehospital and hospital settings.

Purpose

To provide an accurate, reliable, and cost-effective tool for STEMI detection with the potential to redirect human resources into other clinically relevant tasks and save the need for human resources.

Methods

Database: EKG records obtained from Latin America Telemedicine Infarct Network (Mexico, Colombia, Argentina, and Brazil) from April 2014 to December 2019. Dataset: A total of 11,567 12-lead EKG records of 10-seconds length with sampling frequency of 500 [Hz], including the following balanced classes: unconfirmed and angiographically confirmed STEMI, branch blocks, non-specific ST-T abnormalities, normal and abnormal (200+ CPT codes, excluding the ones included in other classes). The label of each record was manually checked by cardiologists to ensure precision (Ground truth). Pre-processing: The first and last 250 samples were discarded as they may contain a standardization pulse. An order 5 digital low pass filter with a 35 Hz cut-off was applied. For each record, the mean was subtracted to each individual lead. Classification: The determined classes were STEMI (STEMI in different locations of the myocardium – anterior, inferior and lateral); Not-STEMI (A combination of randomly sampled normal, branch blocks, non-specific ST-T abnormalities and abnormal records – 25% of each subclass). Training & Testing: A 1-D Convolutional Neural Network was trained and tested with a dataset proportion of 90/10; respectively. The last dense layer outputs a probability for each record of being STEMI or Not-STEMI. Additional testing was performed with a subset of the original dataset of angiographically confirmed STEMI.

Results

See Figure Attached – Preliminary STEMI Dataset Accuracy: 96.4%; Sensitivity: 95.3%; Specificity: 97.4% – Confirmed STEMI Dataset: Accuracy: 97.6%; Sensitivity: 98.1%; Specificity: 97.2%.

Conclusions

Our results remain consistent with our previous experience. By further increasing the amount and complexity of the data, the performance of the model improves. Future implementations of this technology in clinical settings look promising, not only in performing swift screening and diagnostic steps but also partaking in complex STEMI management triage.

Funding Acknowledgement

Type of funding source: None

Innovative techniques to construct powerful artificial intelligence algorithms for st-elevation myocardial infarction

Background

With the sudden advent of Artificial Intelligence (AI), incorporation of these technologies into key aspects of our working environment has become an ever so delicate task, especially so when dealing with time-sensitive and potentially lethal scenarios such as ST-Elevation Myocardial Infarction (STEMI) management. By further expanding into our successful experiences with AI-guided algorithms for STEMI detection, we implemented an innovative ensemble method into our methodology as we seek to improve the algorithm's predictive capabilities.

Purpose

Through the ensemble method, we combined two ML techniques to boost our previous experiments' accuracy and reliability.

Methods

Database: EKG records obtained from Latin America Telemedicine Infarct Network (Mexico, Colombia, Argentina, and Brazil) from April 2014 to December 2019. Dataset: Two separate datasets were used to train and test two sets of AI algorithms. The first comprised of 11,567 records and the second 7,286 records, each composed of 12-lead EKG records of 10-second length with sampling frequency of 500 Hz, including the following balanced classes: unconfirmed & angiographically confirmed STEMI (first model); angiographically confirmed STEMI only (second model); and, for both models, we included branch blocks, non-specific ST-T abnormalities, normal, and abnormal (200+ CPT codes, excluding the ones included in other classes). Label per record was manually checked by cardiologists to ensure precision (Ground truth). Pre-processing: First and last 250 samples were discarded to avoid a standardization pulse. An order 5 digital low pass filter with a 35 Hz cut-off was applied. For each record, the mean was subtracted to each individual lead. Classification: The determined classes were STEMI and Not-STEMI (A combination of randomly sampled normal, branch blocks, non-specific ST-T abnormalities and abnormal records – 25% of each subclass). Training & Testing: The last dense layer outputs a probability for each record of being STEMI or Not-STEMI. These probabilities were calculated for each model (Model 1 trained with Complete STEMI dataset and Model 2 trained with confirmed STEMI only dataset) and aggregated using the mean aggregation to generate the final label for each record. A 1-D Convolutional Neural Network was trained and tested with a dataset proportion of 90%/10%; respectively. Results are reported for both testing datasets (Complete and confirmed STEMI only records).

Results

Complete STEMI Dataset: Accuracy: 96.5% Sensitivity: 96.2% Specificity: 96.9% – Confirmed STEMI only Dataset: Accuracy: 98.5% Sensitivity: 98.3% Specificity: 98.6%'

Conclusion(s)

While Model 1 and Model 2 achieved similar performances with promising results on their own, applying a combination of both through the ensemble model exhibits a clear improvement in performance when applied to both datasets. This provides a blueprint for advanced automated STEMI detection through wearable devices.

Funding Acknowledgement

Type of funding source: None

Artificial intelligence methodology: multi-label classification of abnormal EKG records

Background

Our previous experience with Artificial Intelligence (AI)-conducted EKG characterization displayed outstanding results in fast and reliable identification of Normal EKGs within the International Telemedical System (ITMS)'s massive record repository. By expanding the array of recognizable cardiovascular entities, we upgraded our methodology to accurately discriminate an anomaly amongst a highly complex database of EKG records.

Purpose

To present a feasible AI-guided filter that can accurately discriminate and classify Normal and Abnormal EKG records within a multilabeled cardiologist-annotated EKG database.

Methods

ITMS developed and tested the “One Click”' process, a “Normal/Abnormal” EKG assessing AI algorithm, by incorporating it into their digital EKG reading platform where cardiologists continuously report their findings remotely in real time. To ameliorate the diagnostic range of the algorithm, a separate dataset of 121,641 12-lead EKG records was consolidated from the ITMS database from October 2011 to January 2019. Only de-identified data was used. Preprocessing: The first 2s of each short lead and 9s of the long lead were considered. Limb leads I, II and III; and precordial leads V1, V2, V3, and V5 were used. The mean was removed from each lead. AI models/Classification: Two models were created and tested independently based on the method of EKG acquisition (69,852 records transtelephonic [TTP]; 52,259 mobile transmission [MOB]). Each record is categorized into six disjoint classes based on the most common types of cardiac disorders (Low/null co-occurrence pathologies in these datasets were grouped into analogous groups). Training/Testing: Distribution of both sets per transmission type was performed through a greedy algorithm, which identified multiple diagnoses per EKG record and labeled it separately to the corresponding group, ensuring sufficient samples per class. Detailed class distribution is shown below. An inception convolutional neural network was implemented; “Normal” or “Abnormal” labels were assigned to each EKG record independently and were compared to cardiologists' reports; performance indicators were calculated for each model and group.

Results

MOB model accrued an average accuracy of 86.7%; sensitivity of 90.5%; and specificity of 83.9%. TTP model yielded an average accuracy of 77.2%; sensitivity of 91.1%; and specificity of 69.4% (Lower values were attributed to the “Ventricular Complexes” group, which challenged the algorithm by having a smaller ratio of abnormal exams). Detailed results of each training set are shown below.

Conclusion

Providing an effective and reliable multilabel-capable EKG triaging tool remains a challenging but attainable goal. Continuous systematic enhancement of our AI-driven methodology has led us to satisfactory, yet imperfect results which compel us to further study and improve our efforts to provide a trustworthy cardiologist-friendly triage device.

Funding Acknowledgement

Type of funding source: None

Novel wearable sensor device methodology for STEMI detection

Background

Novel STEMI detection tools using wearable Single Lead EKG methodologies demonstrate vast potential in many clinical scenarios. Recent research suggests that smartwatches and other wearable devices can be repositioned to acquire “new” chest leads that have similar, but not equal, waveforms when compared to traditional precordial leads. Throughout our previous research, only Lead I data had been used to train our Machine Learning (ML) models due to a lack of datasets from these “new” leads. We now propose an innovative methodology to tackle these limitations and compare it with our previous experience.

Purpose

To demonstrate that mathematical vector algebra can reliably transform EKG STEMI databases into different, ML-ready datasets useful to train models with entirely new leads, mainly to be used in the development and training of reliable STEMI detection tools.

Methods

Our previous research has demonstrated that the most accurate (91.2%) ML model was achieved through precordial lead 2 (V2). By definition, V2 corresponds to the difference in electric potential between the Wilson Central Terminal (Wt) and the Chest terminal 2 (C2). To obtain the Wt, at least three electrodes must be used (Right Arm [RA], Left Arm [LA], Left Leg [LL]). Due to practical reasons, we discarded this methodology and worked with Lead I instead, which needs only two body contacts (RA, LA), and provides waveforms that are compatible with the majority of wearable devices (smartwatches, rings, among others). New waveforms (Vn') were obtained by positioning a single lead-capable wearable device (Smartwatch) to chest positions Cn (C1, C2,...,C6) and touching a second electrode with a right-hand finger, which corresponds to the difference in electric potential between RA and the correspondent conventional Vn chest position, respectively. Using vector algebra, we observe that Vn' corresponds to the sum of −aVR + Vn. Vector mathematical analysis was performed for 5,783 STEMI (50%) and 5,784 Not-STEMI (50%) EKG dataset, obtaining their corresponding new precordial leads Vn'. Following this, the ML Heart Attack Detector model was trained with 10,410 EKG (90%) and tested utilizing 1,157 (10%) EKG. Performance metrics were calculated for each new Lead and compared with our Prior Data.

Results

A 1:1 correlation was seen between our previous and current experiments, with Lead V2' performing as the best overall lead with 91.2% Accuracy, 89.6% Sensibility, and 92.9% Specificity. Complete information on prior and new data are provided below.

Conclusions

With the use of this new methodology, we overcame the inherent limitations of using our best Lead (V2) in a single lead approach for STEMI screening. Further prospective data is needed to validate this approach, but it provides a promising blueprint for automated STEMI detection and management triage through the use of wearable devices.

Funding Acknowledgement

Type of funding source: None

1096 Morning Wakefulness is Associated with Reduced Suicidal Ideation in a Nationally-Representative US Sample

Introduction

Nocturnal wakefulness is a unique risk factor for suicidal ideation in clinical as well as community samples. Preliminary data suggest that morning wakefulness may also be a protective factor against such thinking. However, these associations have not been explored in a nationally-representative dataset.

Methods

Data were collected from the 2015-2016 wave of the National Health and Nutrition Examination Survey. Participants reported typical bedtimes and waketimes. From these values, wakefulness during the night (00:00 to 05:59), morning (06:00 to 11:59), afternoon (12:00 to 17:59), and evening (18:00 to 23:59) was determined. Suicidal ideation was assessed by a question about “thoughts that you would be better off dead, or thoughts of hurting yourself in some way.” Ordinal logistic regression estimated the association between the number of hours awake at particular times of day and the frequency of suicidal ideation. Additional analyses adjusted for demographic factors and depressed mood.

Results

Out of 5133 respondents with available data, 125 reported suicidal ideation several days a week, 36 reported suicidal ideation more than half the days, and 29 reported suicidal ideation nearly every day. When controlling for demographics, morning wakefulness was associated with reduced frequency of suicidal ideation (OR: 0.69, 95% CI: [0.59,0.8]). Controlling for depressed mood attenuated, but did not eliminate, this association. Nocturnal wakefulness was not associated with suicidal ideation in this sample.

Conclusion

Using data from a nationally representative sample, morning wakefulness was associated with less frequent suicidal ideation. However, previous findings regarding nocturnal wakefulness were not replicated. The limited number of individuals in the sample endorsing both suicidal ideation and nighttime wakefulness may have insufficient power to detect an association.

Support

Dr. Grandner is supported by R01MD011600.

0374 Decline in Habitual Sleep Duration Over 10 Years and Worsening Sleep Disparities: Data From NHIS (2006-2015)

Introduction

Despite claims in the media, evidence that habitual sleep has declined in recent years is scant. Few data sources exist that systematically document sleep duration in a nationally representative sample, in the same way, over several years.

Methods

Data from 10 years of the National Health Interview Survey were used (N=305,555). During all years, habitual sleep duration, age, sex, race/ethnicity, and height/weight were recorded in the same way. Weighted regression analyses examined sleep duration as the outcome, year as linear predictor, and sociodemographics as covariates. Then, interaction terms examined whether the linear change associated with years was differentially experienced by different sociodemographic groups.

Results

The linear trend of sleep duration over the past 10 years is a loss of 0.78 minutes per year (95%CI -0.91,-0.64; p<0.0001). After adjustment for age, sex, race/ethnicity and BMI, this remained relatively unchanged at 0.86 minutes (95%CI -0.99,-0.73; p<0.0001). A year-by-race/ethnicity interaction was observed (p<0.05). In stratified analyses, Non-Hispanic Whites showed a loss of 0.68 minutes per year (95%CI -0.84,-0.52, p<0.0001). This was 1.33 minutes/year in Blacks/African-Americans (95%CI -1.74,-0.92; p<0.0001), 1.57 minutes/year in Mexican-Americans (95%CI -1.98,-1.16; p<0.0001), 0.99 minutes/year in other Hispanics/Latinos (95%CI -1.51,-0.47; p<0.0001), 0.74 minutes/year in Asians (95%CI -1.24,-0.25; p=0.003), and 1.80 minutes/year in American Indians/Alaskan Natives (95%CI -3.57,-0.03, p=0.046).

Conclusion

On average, the US population has lost 47 seconds of nightly sleep per year over a 10-year period, equating to about 4.7 hours of sleep per year, but racial/ethnic groups were impacted differently. Compared to Non-Hispanic Whites, Blacks/African-Americans lost 96% more sleep, Mexicans lost 131% more sleep, other Hispanics/Latinos lost 46% more sleep, Asians lost 9% more sleep, and American Indians lost 165% more sleep. Thus, sleep disparities may be widening.

Support

Dr. Grandner is supported by R01MD011600

0243 Community-Level Daytime Sleepiness and Substance Use: Implications of Sleep Time and Mental Health

Introduction

Daytime sleepiness is associated with impaired functioning and well-being. Those with more sleepiness may turn to illicit substances to overcome these problems. The present study examined whether community-level daytime sleepiness is associated with the likelihood of drug use.

Methods

Data were pulled from the Sleep and Healthy Activity, Diet, Environment, and Socialization (SHADES) study of N=1007 community adults (age 22–60). Daytime sleepiness was assessed with the Epworth Sleepiness Scale (ESS). Use of different substances was assessed with the Alcohol, Smoking, and Substance Involvement Screening Test (ASSIST). The present analyses examined use within the past month of alcohol, tobacco, cannabis, cocaine, amphetamines, inhalants, sedatives, hallucinogens, and illicit opioids. A separate item assessed caffeine. Ordinal logistic regression analyzed ESS score as a predictor of frequency of substance use adjusted for age, sex, education, and race/ethnicity. Additional models included habitual sleep duration and score on the PHQ9 depression scale.

Results

In sociodemographically-adjusted analyses, ESS score was associated with an increased risk of using tobacco (OR=1.04, p=0.015), cannabis (OR=1.04, p=0.014), cocaine (OR=1.07, p=0.009), amphetamines (OR=1.06, p=0.025), inhalants (OR=1.13, p=0.002), sedatives (OR=1.07, p=0.003), hallucinogens (OR=1.12, p=0.001), and opioids (OR=1.12, p=0.0001). Controlling for sleep duration did not significantly affect these relationships, while controlling for depression made every association non-significant except hallucinogens (OR=1.09, p=0.040).

Conclusion

Daytime sleepiness was associated with increased use of nearly all drug categories, but not alcohol or caffeine. Public consumption of alcohol and caffeine might be sufficiently common that the presence of their use cannot be adequately associated with sleepiness. Moreover, the increased frequency of drug use with sleepiness is not linked to sleep deprivation but may reflect emotional distress.

Support

This work was supported by a grant from Jazz Pharmaceuticals

Dr. Grandner is supported by R01MD011600

0232 Impact of Mental Health on 10-Year Trends in Habitual Sleep Duration

Introduction

Public health efforts aimed at reducing the decline in habitual sleep duration have not been successful. It is possible that this decline is differentially experienced relative to individuals’ mental health status. This would further support the need to focus on mental health as a strategy for improving sleep in the general population.

Methods

We examined 10 years of the National Health Interview Survey data (N=305,555). During all years, habitual sleep duration, age, sex, race/ethnicity, and height and weight (used to compute body mass index) were recorded in the same way. In addition, depressed mood in the past 30 days was evaluated (coded as none, mild, moderate, or severe). Weighted regression analyses examined sleep duration as an outcome, year and depressed mood as predictors, and sociodemographics as covariates. A year-by-depressed mood interaction was computed, and analyses were stratified by group.

Results

There was a significant year-by-depression interaction on linear change in sleep duration over the 10 year period (p=0.0001). Analyses were then stratified by depressed mood. In adjusted analyses, individuals with no depressed mood lost an average of 0.68 minutes of sleep per year (95%CI -0.82,-0.55; p<0.0001). Among those with mild depression, this was 7% higher, at 0.73 minutes (95%CI -1.13,-0.33; p<0.0001). Among those with moderate depressed mood, this was 154% higher, at 1.73 minutes lost per year (95%CI -2.31,-1.16; p<0.0001). Among those with severe depressed mood, this was 351% higher, at 3.07 minutes per year (95%CI -4.22,-1.92; p<0.0001).

Conclusion

The 10-year linear decline in habitual sleep duration seems to depend on mental health status. Individuals with better mental health lose less sleep over time, relative to those with worse mental health. This highlights the importance of mental health as a possible avenue for improving sleep health in the population.

Support

Dr. Grandner is supported by R01MD011600

P1709 Left atrial strain: a novel prognostic marker in patients with chronic kidney disease

Funding Acknowledgements

Nil

Background and objectives

Patients with chronic kidney disease (CKD) are at increased risk of adverse cardiovascular events that are underestimated by traditional risk stratification algorithms. We sought to determine clinical and echocardiographic predictors of adverse outcomes in CKD patients.

Methods

Stage 3 and 4 CKD patients without previous cardiac disease underwent a comprehensive transthoracic and stress echocardiogram, with left ventricular (LV) and left atrial (LA) strain analysis. Participants were followed for the primary end point of a composite of all-cause death and major adverse cardiovascular events (MACE). The secondary end point was a composite of cardiovascular death and MACE.

Results

243 patients (male 63%; mean age 59.2 ± 14.4 years) were followed for a median of 3.9 ± 2.7 years. 69 patients met the primary endpoint and 58 the secondary end point. Age (p < 0.01), history of diabetes mellitus (p < 0.01), indexed LV mass(LVMI) (p < 0.01), LV global longitudinal strain(GLS) (p < 0.01), indexed LA volume(LAVI) (p < 0.01), E/e’ ratio (p < 0.01) and LA strain (LAs) (p < 0.01) were independent predictors of death and MACE. On Cox proportional hazards regression analysis, LAs (p < 0.01) was the only independent predictor for the primary end point in a model accounting for age, diabetes mellitus, LVMI, LVGLS, E/e’ and LAVI. LAs remained an independent predictor for the secondary end point.

Conclusions

LAs is an independent predictor of death and MACE in CKD patients, in whom the predominant cardiac abnormality is diastolic impairment. LAs is a prognostic biomarker, reflecting alterations in diastolic function in CKD.

Abstract P1709 Figure. Kaplan Meier curve of LAs

1679 E/e predicts exercise capacity and adverse cardiovascular outcomes in patients with chronic kidney disease

Funding Acknowledgements

Nil

OnBehalf

NA

Background

Patients with chronic kidney disease(CKD) have reduced physical fitness that contributes to the disproportionately elevated risk of cardiovascular disease in this population. Our aim was to assess the association between E/e’ and exercise capacity in CKD patients and the prognostic role of E/e’.

Methods

Patients with Stage 3/4 CKD, without previous cardiac disease were prospectively recruited. Recruited patients underwent transthoracic echocardiogram and exercise stress echocardiogram with assessment of exercise E/e’. Patients were compared, one to one, to age, gender and risk factor matched controls and were followed annually for 5 years for cardiovascular death and major adverse cardiovascular events (MACE). Exercise capacity was assessed as metabolic equivalents (METs) with reduced exercise capacity defined as METS of ≤7. Raised exercise E/e’ was defined as exercise E/average e’ of >13.

Results

156 CKD patients (62.8 ± 10.6 yrs, male 62%) were compared to 156 matched controls. CKD patients had higher rates of anemia (p < 0.01), larger left ventricular indexed mass (p < 0.01), larger LAVI (p < 0.01) and higher resting (p < 0.01) and exercise E/e’ (p < 0.01). Overall, CKD patients achieved lower METs (p < 0.01) with exercise and a greater proportion of CKD patients had METs ≤7 (p < 0.01). Receiver operating curves (Figure1) showed exercise E/e’ (AUC 0.89, CI 0.84-0.95, p < 0.01) to be the strongest predictor of reduced exercise capacity in CKD patients. Exercise E/e’ of >13 was also associated with higher rates of cardiovascular death and MACE amongst CKD patients.

Conclusion

Exercise E/e’ is a strong predictor of exercise capacity amongst CKD patients, who commonly have reduced exercise capacity presumably consequent to diastolic dysfunction. Raised exercise E/e’ in CKD patients is predictor of cardiovascular death and MACE.

Abstract 1679 Figure.

P246 Exercise E/e refines diagnosis of diastolic impairment in patients with chronic kidney disease

Background

Diastolic dysfunction is common in chronic kidney disease (CKD) patients and is associated with reduced exercise capacity. The ratio of early diastolic mitral inflow velocity to early diastolic mitral annulus velocity (E/e’) correlates with invasively measured left ventricular pressures, both at rest and during exercise. The aim of our study was to evaluate the relationship of resting and exercise E/e’ to exercise capacity in patients with stage 3 and 4 CKD.

Methods

CKD patients were prospectively recruited. All patients were required to be in sinus rhythm, without previous cardiac disease and with normal LVEF. The E/e’ ratio was assessed at rest and patients were stratified into 3 groups based on conventional clinical classifications (normal: ≤8, indeterminate: 8-14, raised: >14). Patients then underwent a symptom limited exercise stress echocardiogram whereby patients with resting E/e’ <14 additionally had post exercise E/e’ measured. Exercise capacity was assessed as metabolic equivalents (METs). A raised exercise E/e’ was defined as exercise E/e’ of ≥13. Lastly, we grouped patients based on abnormalities of baseline and exercise E/e’ (resting E/e’<14 & exercise E/e’<13, resting E/e’<14 & exercise E/e ≥13, resting E/e’ ≥14) and the relationship to METs achieved was evaluated.

Results

The cohort consisted of 156 CKD patients (62.8 ± 10.6 yrs, male 62%). 85% had hypertension and 39% had diabetes mellitus. 52/156 (34%) CKD patients had resting E/e’ of ≥14. Of the 104/156 (66%) CKD patients with normal/indeterminate resting E/e’, 66/104(63%) had normal exercise E/e’ <13 whilst 38/104(37%) had exercise E/e’ ≥13. Patients with resting E/e" & raised exercise E/e’ ≥13 achieved significantly lower METs compared to CKD patients with resting E/e’ <14 & exercise E/e’ <13 (5.0 ± 2.2 vs 8.8 ± 2.5, p< 0.01). Of note, METs achieved in this group were comparable to that achieved by CKD patients with raised resting E/e’ ≥ 14 (5.0 ± 2.2 METs vs 5.3 ± 2.6 METs, p = 0.55).

Conclusion

In CKD patients with normal/indeterminate resting E/e’, there is utility in performing exercise E/e’, which was increased in 37% of patients. Raised exercise E/e’ is associated with poorer exercise capacity and is a potential tool to refine diastolic abnormalities in patients with CKD.

Group Total (n,%) METs Achieved Exercise Duration (minutes) Resting E/e’ ≤ 14 & Exercise E/e’ ≤ 13 66/156 (42%) 8.8 ± 2.5 8.4 ± 3.1 Resting E/e’ ≤ 14 & Exercise E/e’ ≥ 13 38/156 (24%) 5.0 ± 2.2 6.0 ± 3.1 Resting E/e’ ≥ 14 52/156 (34%) 5.3 ± 2.6 5.6 ± 2.2

P6422Physician vs machine: an innovative ST-elevation myocardial infarction pathway through artificial intelligence

Background

The diagnosis of ST-Elevation Myocardial Infarction (STEMI) has traditionally relied on a cardiologist's interpretation of an Electrocardiogram (EKG). This cumbersome process is costly, inefficient and out of date. Artificial Intelligence (AI) -guided algorithms can provide point-of-care, accurate STEMI diagnosis that will facilitate STEMI management.

Purpose

To demonstrate the feasibility of an automated AI-guided EKG analysis for STEMI diagnosis.

Methods

An observational, retrospective, case-control study. Sample: 8,511 EKG cardiologist-annotated records, including 4,255 STEMI cases. Records excluded patient and medical information. The sample was derived from the private International Telemedical Systems (ITMS) database. LUMENGT-AI Algorithm was employed. Preprocessing: detection of QRS complexes by wavelet system, segmentation of each EKG into individual heartbeats (90,592 total beats) with fixed window of 0.4s to the left and 0.9s to the right of main QRS; Classification: A 1-D convolutional neural network was implemented, “STEMI” and “Not-STEMI” classes were considered for each heartbeat, individual probabilities were aggregated to generate the final label for each record. Training & Testing: 90% and 10% of the sample were used, respectively. Experiments: Intel PC i7 8750H processor at 2.21GHz, 16GB RAM, Windows 10 OS with NVidia GTX 1070 GPU, 8GB RAM.

Results

The model achieved an accuracy of 96.5%, with a sensitivity of 96.3%, and a specificity of 96.8%.

Conclusion(s)

1) AI-guided interpretation of the EKG can reliably diagnose STEMI; 2) AI algorithms can be incorporated into ambulance systems for pre-hospital diagnosis, single page activation, emergency department bypass, facilitating more efficient STEMI pathways.

P1464Adoption of feedback to validate a machine learning model for single lead STEMI detection

Background

We have explored the performance of a single lead EKG with Artificial Intelligence (AI) based algorithms in STEMI diagnosis, thus far lead V2 has yielded the best results. Anticipating the performance of the LUMENGT-AI model, we designed a feedback strategy with healthcare centers to expand the validation of our work.

Purpose

To create a pragmatic alternative to the existing gold standard, a 12-lead EKG, for STEMI diagnosis.

Methods

An observational, retrospective, case-control study. Sample: 2,543 exclusively STEMI (anterior, inferior and lateral wall) diagnosis, EKG records. Feedback: From healthcare centers, confirming STEMI diagnosis and location, was obtained (thrombolysis, primary Percutaneous Coronary Intervention (PCI), pharmaco invasive therapy or coronary artery bypass surgery). Records excluded other patient and medical information. Sample was derived from the private International Telemedical Systems (ITMS) database. LUMENGT-AI Algorithm was employed. Preprocessing: detection of QRS complexes using the wavelet system, segmentation of each EKG into individual heartbeats (53,667 total beats) with fixed window of 0.4s to the left and 0.9s to the right of main QRS; Classification: A 1-D convolutional neural network was implemented; “STEMI” and “Not-STEMI” classes were considered for each heartbeat per lead; individual probabilities were aggregated to generate the final label for each record. Training & Testing: 90% and 10% of the sample were used, respectively. Experiments: Intel PC i7 8750H processor at 2.21GHz, 16GB RAM, Windows 10 OS with NVidia GTX 1070 GPU, 8GB RAM.

Results

V2 was the most precise lead with an Accuracy of 93.6%, a Sensitivity of 89%, and a Specificity of 94.7%.

Conclusions

The strategic adoption of feedback from healthcare centers provided strong validation of our model. The results of AI-augmented, single lead EKG are encouraging. We anticipate that this approach will become a promising methodology in STEMI detection.

P1524Women are from venus and face barriers for AMI treatment

Background

The Latin American Telemedicine Infarct Network (LATIN), a program for managing population-based AMI care, has utilized an innovative hub-spoke strategy and remote guidance to expand medical access in remote parts of Brazil, Colombia, Mexico, and Argentina. Based on the Global Lumen Organization for Women (GLOW) project, that has previously demonstrated gender inequalities and worse outcomes for women who undergo Primary PCI, our research explored the management of female patients in our network.

Purpose

To demonstrate gender disparities in a telemedicine-guided, population based AMI program.

Methods

Meta-analysis of >750,000 patients within LATIN. Telemedicine was employed as a screening tool to accurately diagnose AMI. According to the duration of chest pain and transfer time, AMI patients were triaged into guidelines-based pathways of thrombolysis, pharmaco-invasive management or Primary PCI. Resource allocation was identical for men and women.

Results

Data from 784,395 screened patients reveals broad gender disparities. The critical results yielded that female patients have a lower prevalence of STEMI diagnosis and treatment regardless of the technique – PCI, pharmaco-invasive or CABG (p<0.01). These findings assume more significance given the fact that more women than men, were screened through the Telemedicine pathway.

Conclusions

LATIN gender data signals the urgency to demand better AMI care for women.

P6421Can cardiologists rely on artificial intelligence to identify the culprit vessel in STEMI?

Background

The importance of culprit lesion identification is critical for risk stratification of a patient with an ST-Elevation Myocardial Infarction (STEMI). The aforementioned provide patients with a more elaborated strategy of management and treatment either they are treated with PCI or less invasive techniques such as thrombolysis. We report a novel approach that employs AI-guided electrocardiogram (EKG) algorithms for rapid and accurate identification of the culprit STEMI vessel.

Purpose

To create an innovative, machine learning tool for a more effective risk stratification of STEMI patients.

Methods

An observational, retrospective, case-control study. Sample: 2,542 exclusively STEMI diagnosis EKG records that included post discharge feedback from healthcare centers, confirming diagnosis and culprit vessel (Left Main Coronary Artery [LMCA]; Left Anterior Descending [LAD]; Right Coronary Artery [RCA]; Left Circumflex Artery [LCX]; Saphenous Vein Graft [SVG]). Records excluded other patient and medical information. The sample was derived from the private International Telemedical Systems (ITMS) database. LUMENGT-AI Algorithm was employed. Preprocessing: detection of QRS complexes using a wavelet system, segmentation of each EKG into individual heartbeats (27,125 total beats) with fixed window of 0.4s to the left and 0.9s to the right of main QRS; Classification: A 1-D convolutional neural network was implemented; “LCMA”, “LAD”, “LCX”, “RCA”, “SVG”, and “No Information” classes were considered for each heartbeat; individual probabilities were aggregated to generate the final label for each record. Training & Testing: 90% and 10% of the sample was used, respectively. Experiments: Intel PC i7 8750H processor at 2.21GHz, 16GB RAM, Windows 10 OS with NVidia GTX 1070 GPU, 8GB RAM.

Results

Global Accuracy: 79.4%; LAD: Sensitivity 86.2%; Specificity 84.8%. RCA: Sensitivity 85.7%; Specificity 83.7%. LCX: Sensitivity 43.5%; Specificity 96.9%.

Conclusions

Coupling an AI-augmented algorithm and 12-lead EKG provides encouraging results for STEMI culprit vessel localization. Overall, risk stratification is possible for individual lesions located in the LAD and RCA. However, our approach yielded uncertain results in the LCX territory. We plan to continue to exploring variables for improvement of our results.

P5241Balancing limited resources, infra-structure deficits & cultural differences in sustaining the growth of LATIN telemedicine program

Background

In resource-constrained nations, population-based AMI coverage is daunting. Telemedicine can transform with efficient, cost-effective and scalable programs. We present our data with screening >780,000 patients with innovative hub and spoke strategies.

Purpose

Scientifically pristine protocols, rigorous training, unflinching quality assurance, technology upgrades and education of broad stakeholders are essential attributes for creating population-based AMI programs.

Methods

Latin America Telemedicine Infarct Network (LATIN) required methodical groundwork during a 12-month pilot prior to its formal launch and sustenance for 5 years. It involved scrupulous site selection, technology, and telemedicine optimization and system-wide process metrics. Spokes are the LATIN nucleus and require constant (3-T) training: Triage, Telemedicine, and Transportation. Plus, a mandatory deconstruct of their role in LATIN, of urgent transfer and desist non-critical care. Telemedicine requires constant upgrading of platform, tele-equipment and cloud computing. Ambulance availability is a constant challenge as is the battle with payers. Data entry has required meticulous training and oversight. Strict QA processes have monitored critical metrics: Spokes (Door In Door Out, DIDO and Transport Times); Hubs (Door to Balloon Times, D2B); Telemedicine Platform (Time to Telemedicine Diagnosis, TTD).

Results

Linear growth is observed in the number of sites and telemedicine screenings with simultaneous and sustained improvements in D2B and TTD. 784,395 patients were screened at 350 LATIN centers (Brazil 143, Colombia 118, Mexico 82, Argentina 7). With expanded reach, 8,440 (1.08%) patients were diagnosed and 3,924 (46.5%) urgently reperfused, including 3,048 (77.7%) with PCI. Time to Telemedicine Diagnosis (TTD) was 3 min, tele-accuracy 98.9%, D2B 51 min and in-hospital morality 5.2%. Major reasons for non-treatment were insurance, lack of ICU beds and delayed presentation.

Conclusions

As other regions of the world develop large, population-based AMI management initiatives, LATIN can provide important lessons in the sustainability of these processes.

P1928Global telemedicine initiatives for combating ami

Background

Major disparities exist between developed and developing countries in Acute Myocardial Infarction (AMI) outcomes. Telemedicine has emerged as a powerful, cost-efficient, and scalable tool for population-based AMI management. We propose efficient telemedicine protocols as frontline AMI strategies for resource-constrained developing countries.

Purpose

To create a global template of using telemedicine protocols for treating AMI.

Methods

A hub and spoke strategy was utilized for Latin America Telemedicine Infarct Network (LATIN) to expand access in Brazil, Colombia, Mexico, and Argentina. Small clinics and primary care health centers in remote areas were strategically connected with 24/7 primary PCI facilities. Experts at 4 remote sites provided urgent EKG diagnosis and tele-consultation that triggered ambulance dispatch and implementation of standardized AMI protocols.

Results

784,947 patients were screened for AMI at 350 LATIN centers (Brazil 143, Colombia 118, Mexico 82, Argentina 7). With this expanded geographic reach, 8,448 (1.08%) patients were diagnosed with STEMI and 3,911 (46.3%) urgently reperfused, including 3,049 (78%) with Primary PCI. Time to Telemedicine Diagnosis (TTD) was 3 min, tele-accuracy 98.9%, D2B 51 min, and in-hospital mortality 5.2%. Major reasons for non-treatment were insurance denials, lack of ICU beds and chest pain >12 hours.

Conclusions

LATIN demonstrates the feasibility of a population-based and telemedicine guided AMI strategy that can hugely expand access. Telemedicine has important public health implications as a global approach to AMI care in developing countries.

P3352Telemedicine transcends national boundaries in quest of creating a behemoth ami program

Background

The Latin America Telemedicine Infarct Network (LATIN) has exploited the remarkable competence of telemedicine for remote guidance. In doing so, LATIN created a mammoth population-based AMI network that employed experts located several hundred miles away to guide the reperfusion strategies for almost 800,000 screened patients. In this pioneering project, telemedicine was initially utilized to guide AMI management within national confines. We speculated whether LATIN telemedicine navigation could outstrip countrywide borders.

Purpose

To maximally harness the vast possibilities of telemedicine for improving AMI care.

Methods

During its pilot phase, LATIN began as a hub and spoke, AMI system in Colombia where 20 spokes (small community health centers and rural clinics) were configured with 3 hubs that could perform Primary PCI. These sites were linked through web-based connectivity. Expert cardiologists, located 50–250 miles away in Bogota, Colombia, used sophisticated telemedicine platforms for urgent EKG diagnosis and teleconsultation of the entire AMI process. Based upon the duration of chest pain and travel time to the hub, these experts guided patients through guideline-based strategies of thrombolysis, pharmaco invasive management or primary PCI. Efficiency of the telemedicine process was measured with the new metric of time to telemedicine diagnosis (TTD). Cloud computing, GPS navigation, and numerous business intelligent tools were gradually incorporated into LATIN telemedicine. As systems became more scalable, the program was expanded to Brazil, where LATIN flourished. Over the last 18 months, LATIN telemedicine capabilities have been pressed across national boundaries. Presently, all 82 LATIN centers in Mexico are guided by experts located in Bogota, Colombia and the 7 Argentina centers channeled through Santiago, Chile.

Results

784,947 patients were screened for AMI at 350 LATIN centers (Brazil 143, Colombia 118, Mexico 82, Argentina 7). Navigation pathways are depicted in the attached figure. TTD remains extremely low in all four countries, and comparable efficiency and tele-accuracy have been achieved. With expanded geographic reach, 8,448 (1.08%) patients were diagnosed with STEMI and 3,911 (46.3%) urgently reperfused, including 3,049 (78%) with Primary PCI. Time to TTD ranged between 2.8 to 5.8 minutes, with a mean of 3.5 min. Tele-accuracy was 98.5%, D2B 51 min, and in-hospital mortality 5.2%. Various other comparative metrics for the 4 countries are being gathered and will be available at the time of presentation.

Conclusions

LATIN demonstrates the robust ability of telemedicine to transcend national boundaries to guide AMI management. This strategy can be adopted in under-developed countries in Asia and Africa to provide an umbrella of AMI care for the millions of disadvantaged patients.

P1742Is time to telemedicine diagnosis (TTD) analogous to door to balloon time?

Background

Telemedicine is a powerful, cost-efficient, and scalable tool for population-based AMI management. Traditional metrics of D2N, D2B do not gauge telemedicine effectiveness. We explored the utility of TTD in 784,947-screened patients within the Latin America Telemedicine Infarct Network (LATIN).

Purpose

To evaluate the competence of TTD as an efficiency indicator in telemedicine.

Methods

LATIN employed a spoke-hub strategy to expand access in Brazil, Colombia, Mexico, and Argentina. Small clinics (spokes) in remote areas were strategically connected to PCI-capable facilities (hubs). Experts at 4 remote locations provided urgent EKG diagnosis via tele-consultation, additionally, they triggered ambulance dispatch and implementation of guidelines-based protocols. Investing in updated telemedicine technology provided a system-wide TTD reduction.

Results

714,450 patients were screened for AMI at 350 LATIN centers (Brazil 143, Colombia 118, Mexico 82, Argentina 7). Within our territories 8,448 (1.08%) patients were diagnosed as STEMI; 3,911 (46.3%) were urgently reperfused, of those 3,049 (78%) underwent Primary PCI. TTD was 3 min, demonstrating 98.9% tele-accuracy. D2B was 51 min; in-hospital mortality 5.2%. We encountered a linear correlation between D2B and TTD. The latter was, also, inversely related to the number of screened patients - both associations are favorable for LATIN.

Conclusions

TTD is an important indicator of telemedicine efficiency. LATIN will continue to explore this value's strength and other important associations.

P1466Can physicians trust a machine learning algorithm to diagnose ST elevation myocardial infarction?

Background

For the past years, the medical field has been taking advantage of the endless possibilities that Artificial Intelligence (AI) provides. Using computer-aided devices that can perform and interpret electrocardiograms (EKG) accurately pushes current healthcare boundaries. We present the LUMENGT-AI, this model can handle large datasets, multiclass diagnoses, complex EKG morphology, and still detect ST Elevation MI (STEMI) accurately.

Purpose

To develop an innovative AI-based system for automated STEMI specific EKG analysis.

Methods

An observational, retrospective, case-control study. Sample: 8,511 EKG records, previously diagnosed as “normal”, “abnormal” (over 200 conditions) or “STEMI” (4,255 cases). Records excluded patient and medical information. The sample was derived from the private International Telemedical Systems (ITMS) database. LUMENGT-AI Algorithm was employed. Preprocessing: detection of QRS complexes by wavelet system, segmentation of each EKG into individual heartbeats (90,592 total beats) with fixed window of 0.4s to the left and 0.9s to the right of main QRS; Classification: A 1-D convolutional neural network was implemented, “STEMI” and “Not-STEMI” classes were considered for each heartbeat, individual probabilities were aggregated to generate the final label for each record. Training & Testing: 90% and 10% of the sample were used, respectively. Experiments: Intel PC i7 8750H processor at 2.21GHz, 16GB RAM, Windows 10 OS with NVidia GTX 1070 GPU, 8GB RAM.

Results

Ground Truth Score – Accuracy (94.1%), Sensitivity (87.8%), Specificity (98.1%) – see the comparison to published data in Table.

Conclusions

A statistical analysis allowed us to compare STEMI recognition efficiency between physicians and our model. The LUMENGT algorithm results secured its place as a reliable tool to diagnose STEMI faster and more accurately than physicians.

P1751LATIN - A template for effective AMI management in developing countries

Background

In resource-constrained nations, population-based AMI coverage is daunting. Telemedicine can transform the situation through an efficient, cost-effective and scalable program called the Latin America Telemedicine Infarct Network (LATIN). We present our innovative hub-spoke strategy, that has served >780,000 patients.

Purpose

To use telemedicine protocols to demonstrate appropriate access to quality AMI care, encompassing remote areas.

Methods

LATIN required technology and process metrics optimization as well as a scrupulous site selection, during a 12-month pilot. Spokes represent our strategy's nucleus; they consist of small, rural clinics and resource-limited facilities that are connected to PCI-capable hubs. Spokes require constant (3-T) training: Triage, Telemedicine, and Transportation. The latter two categories are the most challenging because they demand constant upgrading.

Results

784,395 patients were screened at 350 LATIN centers (Brazil 143, Colombia 118, Mexico 82, Argentina 7). A total of 8,440 (1.08%) patients were diagnosed with AMI; 3,924 (46.5%) were urgently reperfused including 3,048 (77.7%) who underwent PCI. Globally, Time to Telemedicine Diagnosis (TTD) was 3 min exhibiting 98.9% tele-accuracy, D2B was 51 min, additionally, in-hospital mortality was 5.2%. Major reasons for non-treatment of patients were insurance, lack of ICU beds and delayed presentation.

Conclusions

LATIN is a valuable healthcare system prototype for developing countries. Our hub-spoke strategy focuses on providing adequate AMI management for populations. However, aspects such as ambulance availability, insurance denial and lack of ICU beds must be targeted to improve performance.

P6417Increasing the accuracy of a machine learning algorithm for STEMI diagnosis by incorporating demographic variables

Background

Our previous work demonstrated the diagnostic value of Artificial Intelligence (AI) -driven algorithms for ST-Elevation Myocardial Infarction (STEMI). In the present research, we explore the importance of demographic data inclusion, in order to achieve a more accurate diagnosis.

Purpose

To demonstrate that incorporation of demographic variables into the sample records will augment the accuracy of AI-based protocols for STEMI diagnosis.

Methods

An observational, retrospective, case-control study. Demographic data (age and gender) male/female ratio 1.3, ages 98–18 years was added to the sample records. Sample: 8,511 EKG records, previously diagnosed as normal, abnormal (over 200 conditions) or STEMI. Records excluded other patient and medical information. The sample was derived from the private International Telemedical Systems (ITMS) database. LUMENGT-AI Algorithm was employed. Preprocessing: detection of QRS complexes by wavelet system, segmentation of each EKG into individual heartbeats (90,592 total beats) with fixed window of 0.4s to the left and 0.9s to the right of main QRS; Classification: A 1-D convolutional neural network was implemented, “STEMI” and “Not-STEMI” classes were considered for each heartbeat, individual probabilities were aggregated to generate the final label for each record. Training & Testing: 90% and 10% of the sample was used, respectively. Experiments: Intel PC i7 8750H processor at 2.21GHz, 16GB RAM, Windows 10 OS with Nvidia GTX 1070GPU, 8GB RAM.

Results

The model yielded an accuracy of 97.1%, a sensitivity of 96.8%, and a specificity of 97.5%.

Conclusions

The ability of AI-guided algorithms to diagnose STEMI is increased by expanding the morphological variables with demographic data. This approach may be applied to improve the EKG diagnosis of other cardiovascular entities and improve clinical management.