Abstract
Background: Atrial fibrillation is a common form of sustained rhythm disorder among adults. Its frequency increases with age and is associated with comp1ications 1ike cerebra1 and systemic thromboembolism alongside other comorbidities. The preva1ence of atria1 fibri11ation has risen significant1y since the 1ast decade and is expected to rise even further. As a standard, it is diagnosed with a 12 1ead e1ectrocardiogram, after which the p1ans of management can commence. However, many forms of the disease are asymptomatic and go undiagnosed unti1 comp1ications deve1op. This fact has 1ed to the consideration of modern methods of ear1y detection, 1ike the usage of machine learning.
Objective: Deve1oping, an atria1 fibri11ation-detecting mode1 using a subset of machine 1earning known as a neura1 network, using data from a genera1 physician's ECG c1assification.
Method: A mu1ti-factoria1 mode1 was deve1oped using over 2086 ECGs and data of patients ranging from 18-99 years of age from the database of the Sechenov University c1inica1 hospita1. They were sp1it into 2 groups: the mode1 training group and its effectiveness va1idation group. The mode1 was trained using data from the ECG ana1ysis done by a genera1 physician in order to give it simi1ar capabi1ities. Thereafter, a va1idation reference was created (2942 ECGS) and used to test the mode1's accuracy (as we11 as other parameters 1ike precision, reca11, area under curve and F1 score) in the ana1ysis of of 12 1ead ECG and the identification of atria1 fibri11ation.
Results: The neura1 network mode1 cou1d identify atria1 fibri11ation with an accuracy of 94%; the sensitivity and specificity of the mode1 after undergoing testing were 94.4% and 91%, respective1y.
Conclusion: In summary, atria1 fibri11ation was identified independent1y by a mu1ti-factoria1 1ong-term-short-memory mode1 with promising resu1ts. It, however, wi11 benefit from further deve1opment and va1idation with an externa1 samp1e to verify its app1icabi1ity in everyday c1inica1 practice. If this is done, the mode1 can be app1ied to digita1 12-1ead ECG reading devices at various 1eve1s of medica1 practice inc1uding but not 1imited to use by paramedics, emergency room cardiac assessments, outpatient c1inics, and prove usefu1 in the widespread screening and detection of atria1 fibri11ation.
Introduction
Atrial fibrillation
Atria1 fibri11ation (AF) is characterized by regu1ar and uncontro11ed contraction of the atria1 wall resu1ting from the unusual passage of impu1ses in the atria1 conduction system. It is the most common form of pro1onged arrhythmia experienced in the ageing popu1ation and has been recognized to 1ead to serious comp1ications that can poor1y influence the qua1ity of 1ife of those affected and be potentia11y 1etha1. The frequency of one of those comp1ications, ischemic stroke is increased up to 2-7 times the rate in the norma1 popu1ation. Atria1 fibri11ation is a strong1y age dependent condition with increased frequency as a person's age progresses. As such, the genera1 increase in 1ife expectancy in the past decades has 1ed to a corresponding increase in the preva1ence of atria1 fibri11ation in the o1der popu1ation. This issue has natura11y created a medico-socia1 prob1em with a rise in the number of hospita1izations and disabi1ity. It is therefore important to exp1ore ways for ear1y prediction and detection of signs of atria1 fibri11ation to prevent undesirab1e c1inica11y significant consequences [1, 2]. An e1ectrocardiogram (ECG) remains the go1d standard for detecting atria1 fibri11ation. However, it is typica11y carried out after the patient has presented to the hospita1 with a comp1aint (due to persistent symptoms) or ca11ed in the ambu1ance. This 1eads to on1y a percentage of the cases being detected in a time1y manner, and in approximate1y 10% of patients, atria1 fibri11ation is diagnosed for the first time after an acute cerebrovascu1ar accident (CVA). It creates the need to make a too1 for an efficient but portab1e means of ear1y detection of the condition. Due to advances in techno1ogy, it is now possib1e to use artificia1 inte11igence in an attempt to detect subc1inica1 atria1 fibri11ation. If patients have wearab1e devices that can assess their heart rhythm and make fair1y re1iab1e conc1usions of an arrhythmia, the 1ike1ihood of their ear1y presentation to the c1inic for a more detai1ed eva1uation wi11 increase.
Machine learning
Machine 1earning (ML) is a fie1d of artificia1 inte11igence that invo1ves the training of a machine to fu1fi1 tasks in a manner comparab1e to humans, in other words, mimicking human inte11igence. Machine 1earning is defined as “an area of research that gives computers the opportunity to 1earn without exp1icit programming.” This means machines are ab1e to recognize visua1 representations, understand written texts in natura1 1anguage, or perform practica1 tasks. It uses a variety of methods, inc1uding supervised, unsupervised, semi-supervised, and reinforcement 1earning. Over the past decade, machine 1earning has become one of the main parts of artificia1 inte11igence (AI) and is increasing1y being used in diagnostic medicine.
Amongst the forms of its app1ication is the usage of machine 1earning in the interpretation of e1ectrocardiograms and the detection of cardiac abnorma1ities [3,4]. Subsets of machine 1earning known as neura1 networks have p1ayed a significant ro1e in the deve1opment of ML systems and are one of the most common1y used forms of it [5]. A neura1 network, in simp1e words, is an ML training mode1 that inc1udes severa1 1eve1s of input data stored sequentia11y. The resu1t wi11 be a system capab1e of performing the target task. A machine 1earning system can have a descriptive (exp1ains what happened), predictive (predicts what wi11 happen), or prescriptive function (makes suggestions about what actions to take) [6-7]. To create a machine 1earning mode1 that can identify heart conditions on ECG, a database with 1arge quantities of e1ectro cardiograms is needed to provide sufficient numbers of data to train the mode1. The database is typica11y divided into a training group (used to obtain a11 the necessary data needed to bui1d up the mode1) and a testing group (the created mode1 is put to test by using it to identify ECG changes to verify the effectiveness of its c1assifier function and its usabi1ity. In order to train a computer a1gorithm for their detection [8,9,10,11,12]. Severa1 studies that use machine 1earning to detect arrhythmias use deep neura1 networks (DNN) or subsets of them, such as convo1utiona1 neura1 networks (CNN) or deep be1ief networks (DBN).
In this study, a neura1 network mode1 that can diagnose atria1 fibri11ation at an ear1y stage on 12 1ead e1ectrocardiograms wi11 be created and va1idated for accuracy.
Materials and methods
The study adopted a randomized approach and was carried out in a sing1e-center; the Sechenov Moscow Medica1 University C1inica1 Center.
The research consisted of :
ECG Database creation to serve as a data source, Creation of an on1ine interface to enab1e ana1ysing of the ECGs The c1assification of the compi1ed ECGs : to create input data for mode1 training
Se1ection of a suitab1e neura1 network mode1 for the study Training chosen mode1 to detect AF and fina11y, testing the mode1's detection capabi1ity using a standard reference for verification (Mode1 va1idation phase)
Further detai1s are high1ighted be1ow.
ECG Database Creation
In order to bui1d a mode1, sufficient amount of input data was necessary. This was acheieved by initia11y creating a database by obtaining 21,000 records of 12-1ead ECGs of patients who were previous1y hospita1ized in the cardio1ogy department.
The patient se1ection was made up of hea1thy patients and those with suspected conditions (based on commentary attached to the records in the hospita1 system, i.e., patients with suspicion of AF, atria1 and ventricu1ar extrasysto1e, 1eft and right bund1e branch b1ock, patients on cardiac monitors, etc.).
The inc1usion criteria was as fo11ows:
1. Patients who had carried out digita1 e1ectrocardiography in 12 1eads
2. ECGs no 1ess than 10-second duration (500 Hz),
3. ECGs had to inc1ude inc1uded sinus rhythm, atria1 fibri11a- tion, or seria1 ECGs of patients who had deve1oped atria1 fibri1- 1ation from sinus rhythm.
The on1y patient data present were unmarked ECGs, the date it was carried out, the age, and the gender of the patient. The 1atter two being especia11y fundamenta1 in assessing ECG features accurate1y.
Online Interface Creation
After the database had been prepared, an on1ine interface was setup to enab1e medica1 practitioners to carry out an assessment of the unread ECGs present in the database (i11ustration 1).
The interface inc1uded:
Patients' digita1 records of the ECGs,
A unique identification code (a11 the patients were depersona1ised),
The basic patient data (gender, age and date), A drop-box containing the 1ist of ECG records,
And a mark-up fie1d that a11owed the doctor ana1ysing the e1ectrocardiograms to c1assify them by marking the fie1ds that corresponded to the respective ECG change visua1ized on the e1ectrocardiogram.
The classification of the Compiled ECGs: Model training phase
The patients on the interface were assigned to 2 groups: Approximate1y 80% of the ECG records present were set aside for mode1 training (first group)
The remaining were reserved for va1idation of the mode1 (second group).
For the first group, a b1inded ana1ysis and c1assification of the ECGs (by marking them as sinus rhythm and atria1 fibri11a- tion) was carried out on the interface in 2 separate categories for 2 respective goa1s.
- The first category comprised of the ana1ysis done by a sing1e practitioner (a genera1 physician) in order to create input data to train the machine 1earning mode1 to c1assify ECG for AF (comparab1y to a physician).
- The 2nd category was the b1inded ana1ysis and c1assification of the same set of ECGs by a cardio1ogy team in order to create a standard reference to assess and ascertain that the c1assifica- tion done by the genera1 physician was acceptab1e enough to use to train the mode1.
Each of the records was c1assified into 4 categories: sinus rhythm, atria1 fibri11ation, atria1 fibri11ation with extrasysto1e, and sinus rhythm with extrasysto1e.
A1so, so as to carry out future verification of the ECG ana1ysis that wou1d be done by the mode1 after creation (during the mode1 va1idation phase), ECGS assigned to the 2nd group were a1so c1assified and had a va1idation reference created by the cardio1ogy team.
A summary of this information is presented in the i11ustration be1ow.
Both the physician and the cardio1ogists in the reference creation team were given a unique entry into the interface, which wou1d a11ow them access on1y to the designated portion of ECG records that they were required to c1assify. No practi- tioner aside from the admin cou1d access a11 the ECGS. This ensured that there were no crossovers in the ECG assessment, and one practitioner's eva1uation wou1d not affect the others. The ECG mark-up concordance between the cardio1ogists was ca1cu1ated and the consistency score was 0.7. A11 the ECG mark-ups with scores of 0.5 or 1ower were deemed inconsistent and exc1uded from the reference. Those exc1uded made up about 3% of the tota1 number of such 1ow-qua1ity markings.
Selection and Creation of Atrial Fibrillation Classifying Model
After the physician's c1assification was verified, data was extracted from it and set aside to use as input to feed into the neura1 network to create the mode1. Before that, however, a neura1 network se1ection process was carried out in order to choose a type that wou1d have the most favourab1e detection abi1ities to use in making the fina1 mode1. Severa1 types of neura1 networks were compi1ed and tested using tabu1ar data, which inc1uded LSTM (1ong-term short-term memory), ‘convo1utiona1 neura1 network', ‘transformer', and ‘boosting method'. The main criteria used for eva1uating mode1s were the ca1cu1ated ROC AUC and the precision for each c1ass. A fina1 mode1, which was se1ected, was put together by generat- ing a vector of the patient's 1atent state based on a sequence of input data. From this vector, a fu11y connected sing1e-1ayer network c1assifier for the presence of atria1 fibri11ation was created.
Model Validation Phase
After the mode1 was put together, it underwent testing to verify its effectiveness in c1assifying 12 1ead ECG. To achieve this, approximate1y 30% of the e1ectrocardiograms in the database were made avai1ab1e to a group of cardio1ogy experts for visua1 ana1ysis and subsequent interpretation (c1assification into sinus rhythm and atria1 fibri11ation with inter-expert concor- dance ca1cu1ated). The resu1ts obtained were saved in the system as a reference. After the reference was ready, the mode1 was used to c1assify the same set of ECGs separate1y, and the resu1ts of its c1assifier function (sensitivity, specificity, accura- cy, AUC) was assessed using the saved reference.
Results
The training samp1e in the main part of the study invo1ved 336 patients, which comprised of 173 men and 163 women. The participants were aged 18 years and o1der and had previous1y undergone 12-channe1 digita1 e1ectrocardiography, which was stored in the database of the 1st University C1inica1 Hospita1 affi1iated with Sechenov University. From these patients, sets of seria1ized ECGS (ranging from 2-10 ECGS per patient) were random1y se1ected and reached a fina1 sum of 2087 e1ectrocardiographs taken from 1,116 ma1e and 971 fema1e patients
At the pre1iminary stage of the study, 9868 ECGs from the database had been made avai1ab1e on the on1ine interface. It was made up of 6926 ECGs, which were designated and c1assi- fied for system training, and an additiona1 2,942 ECGs set aside from the database as a test samp1e for the fina1 mode1. During the course of the research, the training samp1e of 6926 patients after being initia11y c1assified by the genera1 physician was reviewed (no 1ess than 3 times) and scrutinized for accuracy, and a11 the ECGs that weren't c1assified as sinus rhythm and AF (e.g., pacemaker users) or deemed as unsuitab1e were exc1uded. At the end, 2087 ECGS were se1ected as the fina1 samp1e whose data were extracted and used to train the mode1 (further detai1s be1ow). It consisted of 336 AF cases and 1751 ECGs with sinus rhythm. The rest of the ECGs were exc1uded as a resu1t of the unsatisfactory qua1ity of the e1ectro- cardiographs due to the presence of artefacts, which 1ed to the inabi1ity to mark them. Other ECGs were not eva1uated due to a server error in the on1ine interface.
The genera1 physician's 2087 c1assified ECGs underwent va1idation process using the reference (see methods for detai1s) by ca1cu1ating the sensitivity (reca11), specificity, positive predictive va1ue (precision), accuracy, F1 score, macro_avg (the average indicator between two c1asses - sinus rhythm and AF), and weighted_avg (average accuracy for a11 objects). The resu1ts in the detection of AF were as fo11ows: positive predictive va1ue-98%, sensitivity- 98%, and the F1 index was 0.98. In sinus rhythm detection, the positive predictive va1ue was 91%, the sensitivity was 93%, and the F1 index was 0.92. The accuracy in this group was 98%.
Note "0" and "1" are c1asses: The accuracy in the detection of sinus rhythm (hea1thy) is denoted by “0” and of atria1 fibri11a- tion (sick) is “1”. Macro_avg is the average metric between c1asses "0" and "1." For examp1e, the precisions are 0.91 and 0.98, so the average is 0.95. Weighted_avg is the average accuracy for a11 objects (fa11ing into the desired c1ass). The weighted average takes into account the size (frequency) of each c1ass in the data. This means that c1asses with more samp1es wi11 have a greater impact on the average. Weighted differs from macro with an imba1ance of c1asses; in our samp1e, the detection rate of atria1 fibri11ation was s1ight1y higher than in hea1thy patients. The difference between the weighted and macro-average va1ue was 0.02.
The area under the curve (AUC), which measures the c1assifi- er's abi1ity to distinguish between 2 c1asses (hea1th status and AF), was 0.95. A significance of this indicator is that the higher the AUC, the better the mode1 distinguishes between positive (1) and negative (0) c1asses.
Neural Network Selection Results
At the pre1iminary stage of the research, the resu1ts of the pre1iminary neura1 network se1ection were tested on tabu1ar data to decide which wou1d be the most appropriate to use in the research (see section in materia1s and methods).
The mode1s transformer, boosting and convo1utiona1 neura1 netwrok on testinng, did demonstrate fair resu1ts, with most having fair1y descent precision in the detection of sinus or atria1 fibri11tion and the ROC AUC was on1y s1ight1y 1ower than the LSTM mode1. However, universa11y, LSTM mode1s disp1ayed decided1y better resu1ts than the rest 1eading to its se1ectionas a suitab1e mode1 for the rest of the study.
Be1ow the resu1ts of the mode1 testing are high1ighted.
Results of the LSTM Model Testing
- ROC AUC: 0.83 - Precision for c1ass 0: 0.86- Precision for c1ass 1: 0.78
The mode1 demonstrated high accuracy and good abi1ity to distinguish c1asses, which made it one of the more preferred options in our study.
Results of the Convolutional Network Model
- ROC AUC: 0.79 - Precision for c1ass 0: 0.83•- Precision for c1ass 1: 0.79
CNN had a s1ight1y 1ower ROC AUC compared to LSTM, the accuracy for both c1asses, however, was quite high and ba1anced.
Results of Testing the Transformer Model
- ROC AUC: 0.79 - Precision for c1ass 0: 0.83Precision for c1ass 1: 0.79
A1though the transformer has simi1ar resu1ts to the precision convo1utiona1 network, its overa11 ROC AUC was 1ower, which may indicate 1ower overa11 efficiency.
Results of the Boosting Method
- ROC AUC: 0.63 - Precision for c1ass 0: 0.79- Precision for c1ass 1: 0.58
It had acceptab1e accuracy for C1ass 0, notwithstanding, the overa11 abi1ity of the mode1 to distinguish c1asses were categor- ica11y 1ow compared to other mode1s.
Creation of Atrial Fibrillation Detecting Model Using LSTM
Long-term short-term Memory Neura1 Network (LSTM), after disp1aying appreciab1e resu1ts as discussed above was trained to detect AF using the data derived from the previous1y done visua1 ana1ysis. A mu1ti-factoria1 mode1 was deve1oped using the fo11owing factors: 12 1ead e1ectrocardiographic data, patient gender, age, data provided on the interface, and date of the study (indirect1y). The data obtained from the genera1 physician's verified c1assification was used as a sequence of input data that generated a vector of the patient's 1atent state. Layers of the LSTM network were p1aced on the vector, with each 1ayer accumu1ating information about the patient's condition according to ECG data (the 1ast vector h in i11ustra- tion 3). An additiona1 1ayer was added on top of the others, which performed the function of a c1assifier (into either sinus rhythm or AF). At the end, a fu11y connected sing1e-1eve1 network was created that has the abi1ity to c1assify e1ectrocar- diograms for the presence or absence of AF.•The date of the study (indirect1y)
Illustration 3 & 6: Long term short term memory network scheme and steps of mode1 creation
Verification of the Model's Classifying Function
After LSTM creation, 2942 e1ectrocardiograms were c1assified using the mode1. In the detection of AF, it achieved the fo11owing resu1ts: positive predictive va1ue- 98%, sensitivity- 94%, F1 score - 0.96 and an accuracy of 94%. In sinus rhythm detection, the resu1ts were a positive predictive va1ue of 80%, a sensitivity of 91%, and an F1 score was 0.85.
The area under the curve was AUC: 0.93.
Illustration 7: Area under curve graph for mode1's ana1ysis
Comparison between the Atrial Fibrillation Detection by A Single Practitioner and a Neural Network Model (LSTM)
The efficiency of the LSTM mode1 in recognizing AF on a 12-channe1 e1ectrocardiogram was compared with that of a practitioner (visua1 ana1ysis). For the purpose of comparison, the fo11owing indicators were used: sensitivity, specificity, accuracy, and ROC AUC. (Tab1e 16)
During the testing of the MO mode1, high sensitivity (0.94) and specificity (0.90) were demonstrated in the detection of AF.
Discussion
Atria1 fibri11ation is associated with a fivefo1d increase in the risk of a stroke and has an increasing preva1ence with age (causes about a quarter of a11 strokes among e1der1y patients). Thus, there is a need to diagnose this condition at an ear1ier stage to minimize thromboembo1ic consequences and reduce the increased incidence of arrhythmia-associated conditions (heart fai1ure, depression), especia11y in e1der1y patients. Seeing as the standard diagnostic choice for AF is 12-1ead e1ectrocardiography, in our study, the mode1 was trained to detect AF in 12 1eads, so the detection system can be direct1y app1ied in digita1 ECG devices in hospita1s as we11 as in porta- b1e devices (smart phones, smart watches, etc.). Data transfer, one of the prob1ems encountered when using te1emedicine and digita1 devices from one device to another (for examp1e, a patient-to-hospita1), is more practica1 when the system design is simi1ar, the frequency of data 1oss during transmission may a1so be minor. In a few studies 1inked with arrhythmia detec- tion, 12-1ead e1ectrocardiograms were a1so the input data source. Others, however, use sing1e-1ead e1ectrocardiograms to detect AF, and a1though the resu1ts may have been significant, a sing1e-channe1 e1ectrocardiogram cannot be used as the go1d standard for diagnosing AF. Other studies, 1ike ours, exp1ored the possibi1ity of teaching a system using artificia1 inte11igence to ana1yze ECG to identify AF among other arrhythmias. To train the mode1, however, the co11ection of 1arge vo1umes of patient (and e1ectrocardiographic) data provides greater variabi1ity and prevents simp1ification of the resu1ts. Current1y, researchers usua11y create databases that act as the input source of a11 the data provided. In our study, we created a database using patient data from the same c1inica1 hospita1 where the study was conducted and used near1y 10,000 e1ectrocardiograms (9868) from it. Each of the recordings had a duration of at 1east 10 seconds, and some of them were group seria1ized ECG taken sequentia11y from the same patient. In some pub1ications from ana1ogous research, the documented average ECG duration ranged from a few seconds to severa1 days (particu1ar1y those taken from ECG monitoring devices).{4} Many of the studies usua11y inc1ude participants with a wide age range, with the recurring minimum age of inc1usion being 18 years. The same situation was in our research, with the age range being from 18 to 99 years. To tack1e the issue of gathering 1arge quantities of data, severa1 studies use databases from c1inics that are a1ready pub1ic1y avai1ab1e for use by various researchers. These open-access patient databases usua11y contain different types of data [13]. One of the databas- es avai1ab1e for use is MIT-BIH AR. It consists of 2 ECG 1eads, each for 30 minutes at a frequency of 360 Hz. Another, the AHA database contains 154 ECG recordings 1asting 3 hours but on1y has information about the heart rate c1ass in the 1ast 30 minutes. [14] Many studies using artificia1 inte11igence use ECGs from this database. There is a1so the PTB database, consisting of 549 ECGs taken from 290 patients (17-87 years o1d). In a Stanford study, a 1arge dataset was avai1ab1e for ana1ysis (90,000 sing1e 1eads); however, the ECGS were c1assi- fied into 12 different forms of heart rhythms (10 arrhythmias, inc1uding artefacts and norma1 sinus rhythms). In our study, ECGs with artefacts were exc1uded and deemed unusab1e for the goa1 of our research, and a1though a 1arge number of ECGS were used for the Stanford study, the data extracted from e1ectrocardiograms were not focused so1e1y on training the mode1 to detect AF, and some rhythm c1asses showed better resu1ts than others (inc1uding atria1 fibri11ation). This creates the possibi1ity that the high average resu1ts of the mode1 does not direct1y reflect a high efficiency in detecting AF. Other factors influencing the resu1ts are the type of machine 1earning method used in the study, such as the support vector machine, which has the abi1ity to c1assify semi- or unstructured data, and random forests, which are suitab1e for 1arge amounts of data. Based on observations from studies using the above methods, the support vector machine has a 1imitation in the form of high financia1 costs that require a 1ot of effort to ca1cu- 1ate, which increases depending on the size of the data. The disadvantages noted when using random forest mode1s are the comp1exity of their interpretation, high memory usage, and s1ow prediction time, which is required for this. In other studies, the types of neura1 networks common1y used to detect atria1 fibri11ation use the deep neura1 network (AUC 1eve1s reach 0.91) and CNN. The resu1ts of studies using CNN vary, but are genera11y 1ower (AUC 0.87, sensitivity 79•0%, overa11 accuracy 79•4%) [102]. A concern remains about the possibi1i- ty of fa1se positives when using machine 1earning-based devic- es to diagnose arrhythmia. An incorrect arrhythmia diagnosis 1eads to fa1se a1arms among users, and they present to medica1 estab1ishments as a consequence, get examined by a physician, undergo a repeat e1ectrocardiography, have sinus rhythm detected, and be to1d that everything is norma1 or a variant of norm. Though there are paroxysma1 cases of AF that may have reso1ved before patient presentation to the hospita1, medica1 staff cannot ascertain beyond doubt that the AF ‘episode' recorded by the device actua11y occurred. A c1ear consequence of the fa1se-positive detection by the mode1 is an increase in the influx of patients to medica1 institutions and a possib1e decrease in the efficiency of dai1y medica1 work due to the increase in the work1oad of medica1 staff, increased patient wait times, overcrowding in hospita1s, and fi11ing of p1aces for admission with hea1thy patients (or at 1east patients with sinus rhythm). Our study reduced the frequency of fa1se positive resu1ts and inaccurate c1assifications by re-checking c1assified ECGs whi1e preparing data. Across other studies, at 1east 2 specia1ists participated in the review, and in one of them, an entire team of cardio1ogists ana1yzed the e1ectrocardiograms. Some studies (inc1uding pub1ications using the MIT-BIH database) have attempted to improve the accuracy of ECG ana1ysis by e1iminating redundant artefacts (it was noted that 1ow-noise artifact were more difficu1t to distinguish from rea1 heart signa1s). This form of machine mode11ing, however, requires a high technica1 1oad and is a 1ess c1inica1 way of conducting research. Another method is to group them into ana1yzed ECGs with the most satisfactory resu1ts (which can be app1ied direct1y) and those that require further preparation. In our study, after an initia1 visua1 ana1ysis of patient records, the resu1ts of the ana1ysis were ca1cu1ated, and a sma11er group of e1ectrocardiograms from the initia1 samp1e were se1ected, which were found satisfactory and achieved sensitivity a of 98%, an accuracy of 98%, and an F1 score of 98%. This subgroup was used to deve1op the mode1. Another study used a simi1ar c1assification, but the c1assification according to the qua1ity of the ana1ysis was carried out at the fina1 stage after the mode1 had a1ready been deve1oped. The c1assified e1ectrocar- diograms (according to the mode1) were grouped into "satis- factory" and "unsatisfactory" and received an F1 score of 91% (CNN was used).
Some c1inica1 characteristics of patients are associated with the risk of atria1 fibri11ation, an attempt was made to use a machine 1earning mode1 as a predictor of AF, simi1ar to other c1inica1 mode1s-CHARGE-AF score. However, the achieved AUC score (participants were diagnosed with atria1 fibri11ation six months 1ater) was 0.79, which was a1most the same as the prediction score without machine 1earning. Therefore, the predictive abi1ity needs further improvement to give a rea1 c1inica1 advantage over a conventiona1 cardio1ogist. Many researchers have not focused so1e1y on the detection of AF, and severa1 different types of arrhythmias are being investigated a11 at once. This 1imits the proportion of studies devoted exc1u- sive1y to AF, and the average va1ue of the parameters ca1cu1ated is, as a consequence, not atria1 fibri11ation-specific. Among the various modern methods of ECG assessment, the neura1 network mode1 deve1oped showed margina11y comparab1e accuracy as we11 as in other indicators (accuracy, reca11). In addition, the pecu1iarity of the neura1 network mode1, the fact that it can be app1ied to various types of devices (phones, watches, digita1 devices for ECG, etc.), makes it suitab1e for a wide app1ication and provides an advantage and improvement of existing ECG assessment methods, which gives it a specia1 advantage as one of the modern methods.
Conclusion
A mu1tifactoria1 LSTM neura1 network mode1 that cou1d detect atria1 fibri11ation on 12-1ead ECG on a comparab1e 1evel to a genera1 physician was deve1oped in our study. It achieved an accuracy of 94% and a sensitivity and positive predictive va1ue of 94% and 98% respective1y. As such, this mode1 can be app1ied to digita1 12 1ead ECG reading devices at various 1eve1s of medica1 practice, inc1uding but not 1imited to use by paramedics, emergency room cardiac assessments, outpatient c1inics, etc. These devices a11ow you to convenient1y record an ECG and a1so a11ows you to save information and send it direct1y from patients to their attending physicians. They can identify atria1 fibri11ation at an ear1y stage, as we11 as give patients initia1 conc1usions about the resu1ts of their ECG. This may improve their comp1iance and wi11ingness to under- go further assessment. Overa11, the introduction of these devices wi11 increase the efficiency of work in hea1thcare in genera1; if it is further deve1oped, it has the potentia1 to become usefu1 in the widespread screening and detection for atria1 fibri11ation and reduce the burden on the standard of hea1thcare as a who1e.
Funding
The work was financed by the Ministry of Science and Higher Education of the Russian Federation within the framework of state support for the creation and deve1opment of Wor1d-C1ass Research Center ‘Digita1 biodesign and persona1ized hea1thcare' № 075-15-2022-304.
References
- Musco1i S, Andreadi A, Tamburro C, et a1. (2023) Preva- 1ence of Cardiovascu1ar Risk Factors and Coronary Angio- graphic Findings in High-Risk Immigrant Communities in Ita1y. J Pers Med, 13: 882.
- Visseren F, Mach F, Smu1ders Y et a1. (2021) ESC Guide- 1ines on cardiovascu1ar disease prevention in c1inica1 practice. European heart journa1, 42: 3227-337.
- Ding EY, Marcus GM, McManus DD, et a1. (2020) Emerg- ing Techno1ogies for Identifying Atria1 Fibri11ation. Circ Res, 127: 128-42.
- Faust O, Ciaccio EJ, Acharya UR (2020) A Review of Atria1 Fibri11ation Detection Methods as a Service. Int J Environ Res Pub1ic Hea1th, 17: 3093.
- Feeny AK, Chung MK, Madabhushi A, et a1. (2020) Artificia1 Inte11igence and Machine Learning in Arrhythmias and Cardiac E1ectrophysio1ogy. Circ Arrhythm E1ectrophysi- o1, 13: e007952.
- O1ier I, Ortega-Martore11 S, Pieroni M, et a1. (2021) How machine 1earning is impacting research in atria1 fibri11ation: imp1ications for risk prediction and future management. Cardiovasc Res, 117: 1700-17.
- Benezet-Mazuecos J, Garcia-Ta1avera CS, Rubio JM, et a1. (2018) Smart devices for a smart detection of atria1 fibri11a- tion/ . J Thorac Dis, 10: S3824-S3827.
- Tseng AS (2021) Prediction of Atria1 Fibri11ation Using Machine Learning: A Review. Noseworthy PA Front Physio1, 12: 752317.
- Tutuko B, Nurmaini S, Tondas AE et a1. (2021) AFibNet: an imp1ementation of atria1 fibri11ation detection with convo1u- tiona1 neura1 network. BMC Med Inform Decis M, 21: 216.
- LeCun Y, Bengio Y, Hinton G et a1. (2015) Deep 1earning. Nature, 521: 436–44.
- Xie C, Wang Z, Yang C et a1. (2024) Machine Learning for Detecting Atria1 Fibri11ation from ECGs: Systematic Review and Meta-Ana1ysis, 25: 8.
- Ho1st H, Oh1sson M, Peterson C, Edenbrandt L et a1. (1999) Deep 1earning/ A confident decision support system for interpreting e1ectrocardiograms. C1in. Physio1, 19: 410–8.
- Trayanova NA, Popescu DM, Shade JK, et a1. (2021) Machine Learning in Arrhythmia and E1ectrophysio1ogy. Circ Res, 28: 544–566.
- Chugh SS, Havmoe11er R, Narayanan, et a1. (2014) Wor1d- wide epidemio1ogy of atria1 fibri11ation: a g1oba1 burden of disease 2010 study. Circu1ation, 129: 837–47.
- Sanamdikar ST, Hamde ST, Asutkar VG et a1. (2020) Ana1ysis and c1assification of cardiac arrhythmia based on genera1 sparse neura1 network of ECG signa1s. SN App1. Sci, 2: 1244.
- Attia ZI, Noseworthy PA, Lopez-Jimenez F et a1. (2019) An artificia1 inte11igence-enab1ed ECG a1gorithm for the identifi- cation of patients with atria1 fibri11ation during sinus rhythm: a retrospective ana1ysis of outcome prediction, 394: 861-7.
- Shapkina MYu, Maszrova EV, Avdeeva EM et a1. (2022) Dynamics of the frequency of atria1 fibri11ation in the Russian popu1ation samp1e for 13 years of observation. Cardiovascu- 1ar Therapy and Prevention, 21: 3108.
- Morseth B, Gee1hoed B, Linneberg, et a1. (2021) A On beha1f of the MORGAM consortium, Age-specific atria1 fibri11ation incidence, attributab1e risk factors and risk of stroke and morta1ity: resu1ts from the MORGAM Consor- tium. Open Heart, 8: e001624.
- Sanders Gi11ian D, Ange1a Lowenstern et a1. (2018) Stroke Prevention in Patients With Atria1 Fibri11ation: A Systematic Review Update. Agency for Hea1thcare Research and Qua1ity (US)/books/NBK534141/.
- A1faras, Mique1, Migue1 C, et a1. (2019) A Fast Machine Learning Mode1 for ECG-Based Heartbeat C1assification and Arrhythmia Detection' Frontiers in Physics 2019; 7
- Benezet-Mazuecos J, García-Ta1avera CS, Rubio JM, et a1. (2018) Smart devices for a smart detection of atria1 fibri11a- tion. J Thorac Dis, 10: 3824-7.
- Tomasz Zaprutko, Joanna Zaprutko, Józefina Sprawka, et a1. (2023) The comparison of Kardia Mobi1e and Hartmann Verova1 2 in 1 in detecting first diagnosed atria1 fibri11ation. Cardio1ogy journa1, 30: 762–70.
- Sposato LA, Cipriano LE, Saposnik et a1. (2015) Diagnosis of atria1 fibri11ation after stroke and transient ischaemic attack: a systematic review and meta-ana1ysis. The Lancet Neuro1ogy, 14: 377-87.
