Quality-Gated Arrhythmia Decision Support System and Method Using Photoplethysmography Signals

Abstract

The invention's distinguishing feature is not a single AFib classification algorithm or a single model architecture. The invention is a quality-gated PPG-based arrhythmia decision support workflow that extends from the measurement initiation step to clinical reporting, validation, and model updating. In this workflow, a mobile electronic device acquires PPG data via a camera and light source; the signal quality assessment gate running on the device subjects the measurement to an accept / reject decision before sending it to the main classification pipeline or the cloud. If the quality is insufficient, the measurement is rejected without classification, and corrective feedback is given to the user. If the quality is sufficient, the PPG signal is transferred to modules for preprocessing, feature extraction, machine learning-based rhythm classification, confidence score calibration, hybrid rule engine, clinician panel, reporting, PPG-ECG / Holter time matching, and model lifecycle. The system does not present the raw model output as a direct diagnosis; it generates decision support output for arrhythmia, rhythm probability, calibrated confidence score, quality information, and validation guidance. In cases of low confidence, discordance, recurrent suspicious results, or clinically significant alarm conditions, the recording is routed for ECG / Holter validation or clinician review. Model updates are managed through validation, calibration, bias testing, shadow mode, and human confirmation steps, rather than clinician corrections being passed on to automated training without supervision.

Description

1 TARIFF 1.1 Title of invention Quality-Gated Arrhythmia Decision Support System and Method Using Photoplethysmography Signals 1.2 Technical area 5 The invention relates to biomedical signal processing, mobile health technologies, and photoplethysmography (PPG)-based rhythm. evaluation, AI-powered clinical decision support systems, remote patient monitoring, and modeling. It relates to the areas of life cycle traceability of a mobile electronic device. More specifically, the invention concerns a mobile electronic device. (110) The PPG signal obtained with the camera (112) and light source (114) is transmitted to a device that works on the device. Passing the signal quality evaluation through the gate (130), low quality measurements are sent to the remote server, 10 Rejected before being sent to the cloud or master rhythm classifier, preprocessing of appropriate measurements module (140), machine learning classification module (160), calibration / confidence score module (162), Hybrid rule / verification and alarm engine (170), PPG-ECG time matching module (200), clinician It is related to connecting to the panel (180) and the model lifecycle / MLOps module (210) steps. 1.3 State of the art 15 ECG-based devices are used in monitoring atrial fibrillation and other cardiac rhythm disorders. wearable sensors, smartwatches, mobile applications, Holter systems, event recorders, and PPG Measurement approaches based on data collection methods are used. Known solutions include smartphone cameras and external devices. Heart rate can be monitored via electrodes, wearable optical sensors, or cloud-based analysis services. Output regarding rhythm irregularities can be generated. 20 However, some known systems can detect low-quality PPG signals before analysis due to a lack of reliability. Instead of rejecting it in this way, correct the motion artifact in the signal or directly measure the raw data. It focuses on referring to the classification algorithm. Home environment, different device types, variable PPG obtained due to finger pressure, lighting conditions, activity level and user behavior. The quality of the records can vary greatly. This situation leads to misclassification, unnecessary cloud inference 25 these calls can lead to increased data transfer, latency, battery drain, and a risk of false alarms in clinical workflow. She can give birth. In the known state of the art, the on-device signal quality evaluation gate (130) is low quality. Before sending the measurement to the remote server, cloud, or master rhythm classifier, a decision must be made to accept / reject it. subjected to; calibrate confidence score of appropriate measurement, hybrid rule / verification and alarm engine (170), PPG-30 ECG time matching module (200), clinician panel (180), asynchronous explainability module (220) and end-to-end technical clinical decision with controlled model life cycle / MLOps module (210). The integrated structure that connects to the support workflow is not adequately presented. 1.4 Technical problem The technical problem that the invention aims to solve is the camera (112) and light of the mobile electronic device (110). Reliable, explainable PPG time series obtained under variable conditions with source (114), To produce a calibrated arrhythmia decision support output that can be integrated into clinical workflows; low transferring high-quality measurements to a remote server, the cloud, or the main rhythm classifier, and identifying faulty rhythms. to prevent classification; low reliability or inconsistent records to ECG validation and referral to clinician review; alarm, reporting, data lifecycle and model update 40 The goal is to manage these processes in a traceable and controlled manner. 1.5 The purpose of the invention The main purpose of the invention is to obtain a mobile electronic device (110) with a camera (112) and a light source (114). The security of the obtained PPG signal is enhanced by the on-device signal quality evaluation gate (130), It can be verified by the clinician, time-matched with ECG recordings, and the model life cycle is 45. The aim is to provide a traceable arrhythmia decision support workflow. Other aims of the invention are given below: • During PPG measurement, contact, light, motion, frame loss, signal amplitude, signal-to-noise ratio, and pulse are measured. Quality index calculator based on at least two of the following indicators: range consistency and signal shape. Calculating the quality index with (132). 2 • Records with an insufficient quality index can be transferred to a remote server, the cloud, or using machine learning. Before sending to the classification module (160), reject on the device and return to the user To give corrective instructions via notification module (134). • The signal of suitable quality is processed through 5 steps of the preprocessing module (140) and feature extraction module (150). by passing it through to prepare for rhythm classification. • The machine learning output should be a calibration / confidence score module, not a raw probability value. (162) Transferring the calibrated confidence score to the clinical decision support workflow. • Classification result based on heart rate, symptom reporting (230), repeated suspicious measurements, quality Hybrid rule / validation and alarm engine (170) 10 which evaluates together with index and confidence score to ensure. • Where necessary, the PPG segment can be timestamped with the ECG reference recording / segment (202). To provide clinical validation, model labeling, and conflict review by matching. • Without displaying the explainability output to the patient interface (182) in a way that gives the impression of a definitive diagnosis, A limited review layer with asynchronous explainability module (220) and clinician panel (180) 15 to present as. • Clinician corrections, reporting, database and audit log module (190), data set version, model Controlled model update by associating the release, calibration result, and deployment version. to evaluate within the cycle. • Raw PPG recordings that do not trigger alarms and recordings requiring clinical review are separated by data lifecycle 20 To ensure data minimization and technical security by subjecting data to cycles. 1.6 Summary of the invention The distinguishing aspect of the invention is that it does not involve a single AFib classification algorithm or a single model architecture. The invention covers everything from the measurement initiation step to clinical reporting, validation, and model updating. This is a quality-gated PPG-based arrhythmia decision support workflow. In this workflow, mobile electronics 25 The device (110) obtains PPG data through the camera (112) and light source (114); on the device working signal quality assessment gate (130), measurement to main classification line or cloud It is subject to an acceptance / rejection decision before shipment. If the quality is insufficient, the measurement is not classified. It is rejected and corrective feedback is given to the user via the user feedback module (134). Quality When sufficient, PPG signal preprocessing module (140), feature extraction module (150), machine 30 learning classification module (160), calibration / confidence score module (162), hybrid rule / validation and alarm engine (170), clinician panel (180), reporting, database and audit log module (190), PPG- ECG time matching module (200) and model lifecycle / MLOps module (210) steps is transferred. The system does not present the raw model output as a direct diagnosis; the decision support output regarding arrhythmia is rhythm 35 The probability generates a calibration confidence score, quality information, and verification guidance. Low confidence, In cases of discrepancy, recurrent suspicious results, or clinically significant alarming conditions, the recording is an ECG. It is referred for verification or clinician review. Model updates, clinician Instead of transferring corrections to automated training without control, validation, calibration, and bias corrections are used. The test is managed through shadow mode and human approval steps. 40 1.7 Technical effects and advantages provided by the invention The main technical effect of the on-device signal quality evaluation gate (130) is to detect low-quality PPG recordings. This is the rejection of the data before sending it to the remote server, the cloud, or the main classifier. This is early rejection. This mechanism not only reduces the risk of clinical error; it also eliminates unnecessary inference calls from the processor. also includes load, battery consumption, data transfer volume, network latency, and cloud computing cost. It restricts. • Poor quality PPG recordings are rejected before analysis, thus supporting false positive and false negative decision support. The risk of failure is reduced. • Instead of trying to correct the motion artifact, the approach of canceling the measurement if the quality is inadequate is different. It provides a traceable acceptance / rejection logic in the device and terms of use. 50 3 • Calibration / confidence score module (162), raw model probability uncontrolled into clinical workflow prevents the transfer of low-security records and directing them to the clinician panel (180). It makes it easier. • Stopping model deployment when calibration fails can lead to erroneous confidence scores in clinical practice. 5 It acts as a technical security lock that prevents access to the decision support interface. • Separation of the asynchronous explainability module (220) from the main classification response time, clinician It provides retrospective review and decision support in its panel (180). • PPG-ECG time matching module (200), clinical validation, discrepancy detection, model labeling It supports the review of production and confidence score. 10 • Reporting, database and audit log module (190), regulatory technical file, quality management and The model enhances lifecycle traceability. • Deleting raw data that did not trigger an alarm from temporary storage and requiring clinical review. Storing data in encrypted form transforms the data lifecycle into a technical security element. 1.8 Brief description of the images 15 • Figure 1 shows the overall system architecture with device-side and server / clinic-side separated, and data between modules. It shows the flow. • Figure 2 shows the low-quality measurement of the signal quality evaluation gate (130) operating on the device. The quality-gated measurement stream that it rejects without sending to a remote server, the cloud, or the master classifier. It shows. 20 • Figure 3, PPG signal preprocessing module (140), feature extraction module (150), machine learning It shows the classification module (160) and calibration / confidence score module (162) line. • Figure 4 shows the timestamp and tolerance of the PPG segment and the ECG reference recording / segment (202). matching across the range and linking to clinical validation and conflict review It shows. 25 • Figure 5, classification result of hybrid rule / verification and alarm engine (170), quality index, together with the confidence score and symptom reporting (230); patient interface (182) and The clinician panel (180) shows the distinction. • Figure 6, reporting, database and audit log module of clinician corrections (190), calibration / confidence score module (162), machine learning classification module (160) and model 30 Connecting to the controlled model lifecycle via the lifecycle / MLOps module (210) It shows. 1.9 List of reference markers Reference mark Technical element 100 General arrhythmia decision support system 110 Mobile electronic devices 112 Cameras 114 Light sources 116 Accelerometers / motion sensors 118 Local memory / temporary storage 120 PPG signal acquisition modules or PPG time series 130 On-device signal quality assessment gate / SQA Gatekeeper 132 Quality Index Calculator 134 User feedback module 140 Preprocessing modules 150 Feature Extraction Module 160 Machine learning classification modules 162 Calibration / confidence score module 170 Hybrid rule / verification and alarm engine 180 Clinician panel 4 182 Patient Interface 190 Reporting, database and audit log module 200 PPG-ECG time matching module 202 ECG reference records / segments 204 Verification label / dispute information 210 Model lifecycle / MLOps module 220 Asynchronous explainability module 230 Symptom reports 1.10 Detailed description of the invention 1.10.1 General system structure General arrhythmia decision support system (100), a mobile electronic device (110), camera (112), light source (114), preferably accelerometer or motion sensor (116), local memory or temporary storage (118), PPG 5 signal acquisition module (120), on-device signal quality evaluation gate (130), quality index calculator (132), user feedback module (134), preprocessing module (140), feature extraction module (150), machine learning classification module (160), calibration / confidence score module (162), Hybrid rule / validation and alarm engine (170), clinician panel (180), patient interface (182), reporting, database and audit log module (190), PPG-ECG time matching module (200), ECG reference 10 record / segment (202), validation tag / discrepancy information (204), model lifecycle / MLOps module (210), asynchronous explainability module (220) and symptom reporting (230) elements at least It may include some of them. It is not necessary for all modules to be located in a single physical device. Device-based signal quality. While the evaluation gate (130) is run on the mobile electronic device (110), preprocessing module 15 (140), machine learning classification module (160), calibration / confidence score module (162), reporting, database and audit log module (190), clinician panel (180) or model life Part of the cycles / MLOps module (210) is on the mobile device, part is on a remote server or cloud. This can be implemented in its infrastructure. The essential technical core is the quality acceptance / rejection of the PPG record on the device. It is subject to a decision and, if the quality is insufficient, it is not transferred to the main classification line. 20 1.10.2 PPG signal acquisition The user’s fingertip through a mobile electronic device (110), camera (112) and light source (114) or obtains an image / frame series from an equivalent tissue region. From the resulting frame series, red, PPG time by subtracting intensity changes related to green, blue or other suitable color channels The series or PPG signal acquisition module output (120) is generated. Measurement time, sampling rate, camera 25 frame rate, device type, operating system, timestamp, patient / user profile, symptom reporting (230) And the measurement context can be stored as metadata. PPG signal acquisition, user-initiated measurement, initiated during symptom reporting (230) The measurement can be carried out in the form of a measurement initiated by a measurement or a scheduled reminder. Mobile during the measurement. electronic device (110), finger contact, excessive pressure, light saturation, camera stability and motion 30 It can generate real-time or post-measurement quality indicators about the level. 1.10.3 On-device signal quality assessment gate / SQA Gatekeeper On-device signal quality evaluation gate (130), rhythm classification of PPG time series It evaluates whether it is suitable before being sent to the main classification pipeline or the cloud. Quality index calculator (132); accelerometer / motion sensor (116) data, finger touch, light saturation, 35 dropped square rate, signal amplitude, signal-to-noise ratio, pulse interval coherence, signal shape stability, It can use dispersion measures such as skewness / kurtosis, or a weighted / regular combination thereof. Quality index, pre-determined or clinical validation, device matrix, age / population characteristics or compared with thresholds that can be updated according to the use case. When quality is insufficient. The system rejects the measurement without sending it to the master rhythm classifier or cloud inference service; 40 This results in unnecessary data transfer, cloud extraction costs, network latency, CPU load, and battery consumption. is reduced. The lack of quality is not to correct the motion artifact, but to determine the acceptance or rejection decision of the measurement. It is used to give. In at least one application example, the on-device signal quality evaluation gate (130) is used in mobile electronics. quantized device (110) that can work offline, has low memory and processor requirements It can be implemented as a small model or a rule-based decision engine. This application example is standalone. The requirements are not limited by a specific model type, model size, or numerical threshold. 5 1.10.4 User feedback and quality rejection When the quality index falls below the threshold, the user feedback module (134) requires the hand to be held still. adjusting finger contact or finger pressure, correcting lighting conditions, camera lens It generates corrective instructions, such as changing the shutdown method or restarting the measurement. The system can terminate the measurement session when the number of quality rejections reaches a certain number, and rejection 10 Reporting the reason, timestamp, device information, and quality criteria, database, and audit log. can store within the scope of module (190). 1.10.5 Pre-processing line Pre-processing module (140) resampling on the PPG signal passing through the quality gate, basal slip removal, high-pass filtration, low-pass filtration, band-pass filtration, polyphase 15 resampling, Butterworth or equivalent filtering, Hampel filter or equivalent outlier detection. reduction, windowing, Z-score or equivalent normalization, smoothing, and pulse peak decrement. It can apply at least one of the following determination procedures. These procedures include measurement time, sampling rate, and device. It can be adjusted according to the type and patient population. For example, 30-second windows, a specific resampling frequency, filter cutoff frequencies 20 or quality control distribution metrics can be used; however, these numerical values ​​are for application examples. It is of a qualitative nature and should not be considered a mandatory element that would narrow the scope of the request. 1.10.6 Feature Extraction Feature extraction module (150), pulse intervals, pulse waveform from preprocessed PPG signal. It generates features, quality statistics, and irregularity measures. In at least one application example, pNN50, 25 SDNN, Poincare SD1 / SD2 ratios, interval irregularity, pulse wave amplitude variability, ascent and descent slope, pulse morphology consistency, pattern irregularity, and intra-measurement quality. Statistics can be extracted. The feature set is not narrowed down to a specific mathematical criterion; signal and clinical It can be expanded or updated based on validation results. 1.10.7 Machine learning classification and model adaptation 30 Machine learning classification module (160), PPG time series, feature vector or these Using a combination of factors, we can consider the possibility of atrial fibrillation, the possibility of normal sinus rhythm, and other rhythm disorders. the possibility of a noisy / uninterpretable measurement label or a combination of these relating to rhythm It produces a classification output. This classification output is not a definitive diagnosis, but rather a clinical decision support output. It is produced as follows: 35 Classification in at least one application example uses a one-dimensional convolutional neural network, 1D-ResNet. temporal convolutional network, residual linked neural network, attention-based layer, decision tree, support vector This can be achieved with the machine or hybrid combinations thereof. The model is supported by open PPG datasets. They may undergo preliminary training; fine-tuning or clinical adaptation with prospective PPG-ECG recordings. It is possible to use layer freezing, partial freezing, low learning rate retraining, and TCN 40. Retraining the header or methods like Elastic Weight Consolidation can improve previous representations. It can be used as an example of a practice for protection and reduction of the risk of catastrophic forgetting. 1.10.8 Calibration confidence score and calibration safety lock Calibration / confidence score module (162) converts raw model output to Platt scaling, isotonic regression, temperature Converts to a calibrated confidence score using scaling or an equivalent probability calibration method. 45 Uncalibrated raw probability values ​​are fed directly into the clinical decision support interface as if they were the final decision. Not presented. Trust score; in clinician panel (180), hybrid rule / verification and alarm engine (170), It is used in the PPG-ECG time matching module (200) and in low confidence recording marking. In at least one application instance, calibration is rerun after each model update. If the calibration success criterion is not met, the model distribution line is stopped or the relevant model version is 50. 6 It is not included in the clinical workflow. Calibration result, validation report, model version and dataset version. It is stored in association with the calibration / confidence score module (162). Thus, the calibration / confidence score module only produces the score. It's not a step, it's a technical security measure that prevents erroneous confidence scores from entering clinical decision support systems. It functions as a lock. 5 1.10.9 Hybrid rule engine and alarm management Hybrid rule / validation and alarm engine (170), machine learning classification module (160) does not use the result alone; heart rate, confidence score, quality index, symptom reporting (230), measurement signal parameters such as time, number of repeated suspicious results, pNN50, SDNN, Poincare SD1 / SD2 and at least one of the threshold values ​​determined by clinical validation, along with the classification result. 10 It evaluates thresholds not as fixed and unchanging values, but based on clinical validation, age, device type, can be determined or updated based on population characteristics or performance monitoring results. They are considered as parameters. Hybrid rule / verification and alarm engine (170), repeated suspicious result in a specific time interval, Low confidence interval, symptom-associated arrhythmia, pre-defined low / high heart rate 15 Limited alert on patient interface (182) when rate threshold or PPG-ECG discrepancy occurs, clinician in the panel (180) priority registration, ECG verification guidance or reporting, database and audit The log module (190) can start recording. 1.10.10 PPG-ECG time matching PPG-ECG time matching module (200), PPG segment T1-T2 time interval, simultaneous or 20 Matches the recent ECG reference recording / segment (202) with the T1±Δ and T2±Δ tolerance range. Matching, device time, server time, measurement start time, data acquisition time or time This can be done using correction coefficients. In at least one application example, correlation over time Matching reliability can be supported by using tolerance or signal matching criteria; these values Fixed mandatory thresholds are not placed in independent demand. 25 Matching PPG and ECG segments, validation label / discrepancy information (204), clinical validation, The model fine-tuning label (PPG) is used for discrepancy detection and confidence score review. When there is a discrepancy between the classification result and the ECG reference label, or when confidence is lacking When the score is low, the relevant record is referred to the clinician panel (180) for review. 1.10.11 Clinician panel, reporting and asynchronous explainability 30 Clinician panel (180), last measurement time, rhythm classification, heart rate for assigned patients, symptom reporting (230), quality status, confidence score, PPG waveform, ECG reference recording / segment (202) can view clinical notes, alarm history and audit log records. Reporting, database and audit log module (190), patient-based reports, alarm logs, access logs, clinician corrections, It can store data export and model update records. 35 Asynchronous explainability module (220), TCN Attention Rollout, SHAP or equivalent explainability It can calculate its output asynchronously in a way that does not increase the main classification response time. This Explainability outputs are presented in the patient interface in a way that gives the user the impression of a definitive diagnosis (182) not shown; decision support, retrospective review, model behavior in clinician panel (180) It is presented as an evaluation and technical file supporting output. Thus, the patient interface (182) and 40 Information presentation and technical isolation in terms of user safety between clinician panel (180). It is provided. 1.10.12 Data lifecycle, data minimization and security Local memory / temporary storage (118) and reporting, database and audit log module (190), raw PPG data, processed data, quality rejection reason, alarm log, clinical validation data and training candidate data 45 It can be configured to handle different data lifecycles. Alarm-untriggered and clinical Raw PPG recording that does not require review, stored on-device temporary storage for a predetermined period. It can eventually be automatically deleted. Alarm, low confidence, discrepancy, or the need for clinical review. When it occurs, the relevant record can be stored in an encrypted database. 7 Measurement initiation, measurement rejection, reason for rejection, classification result, confidence score, alarm generation, clinician viewing, adding clinical notes, ECG matching, model updating, data export, and At least one of the deletion operations is via the reporting, database and audit log module (190) audit log It can be saved as such. Role-based access control, session timeout, encrypted storage, and user 5 Authorization level restrictions are among the viable security options for the system. 1.10.13 Model lifecycle and controlled update Model lifecycle / MLOps module (210), training dataset version, source code version, model weights, hyperparameters, calibration result, validation result, bias test result, shadow mode It associates the comparison result and distribution version with traceable logs. Clinician corrections, 10 The changes are not directly and uncontrollably transferred to the model update. First, the changes are logged. It is verified and, where necessary, marked as candidate training data. Model update candidate workflow includes data retrieval, preprocessing, model training or fine-tuning, and calibration. validation, bias control, shadow mode comparison, clinical and engineering approval, limited distribution and It may include at least two of the drift / performance monitoring steps. Calibration success criterion or 15 If the validation criteria are not met, the relevant model version is not put into production. This structure is part of the model lifecycle. by transforming the cycle from being merely a regulatory compliance element into a technical safety and control layer. It transforms into... 1.10.14 Application examples Example of routine home measurement: The user initiates the measurement on the patient interface (182). The system uses camera (112) and 20 activates the light source (114), PPG time series or PPG signal acquisition module output (120) It creates and makes a quality accept / reject decision with the on-device signal quality evaluation gate (130). Quality If insufficient, the measurement is rejected before being sent to the cloud and the user feedback module (134) The user is given instructions to measure again. If the quality is satisfactory, the signal preprocessing module (140) and the machine The learning is transferred to the classification module (160) line; decision support output and confidence score are generated. 25 Example of measurement during symptom: The user experiences palpitations, dizziness, shortness of breath, chest pain, or The measurement is initiated by selecting one of the similar symptoms as a symptom report (230). Symptom The report (230) is associated with the measurement metadata. The confidence score is low, the probability of arrhythmia is high or If there is a recurring suspicious result, a priority investigation record is created in the clinician panel (180). Clinical validation example: ECG reference 30 with PPG recording at a clinical center or research phase. record / segment (202) PPG-ECG time matching module in the same or near timestamp interval (200) is used to match the model's PPG-based result with the ECG reference label if it is not compatible. or if the confidence score is low, the record is transferred to the clinician panel (180). Clinician correction model training The candidate tag is saved to the set; however, uncontrolled automated retraining is not performed.

Claims

8 REQUESTS 1. Quality gated by a mobile electronic device (110) and / or remote server infrastructure PPG-based arrhythmia decision support method; mobile electronic device (110) camera (112) and light obtaining raw images or frame series from the user's tissue region via source (114); word 5 The subject is the raw image or frame series from which the PPG time series or PPG signal acquisition module output is obtained. (120) creation; PPG time series signal quality working on mobile electronic device (110) evaluation gate (130) by motion, contact, light, frame loss, signal amplitude, signal-to-noise Quality assessment is evaluated according to at least two of the following criteria: rate or shot consistency. If the result does not meet a predetermined or updatable quality threshold, the measurement will be considered invalid. Rejection before being sent to the cloud or master rhythm classifier and user feedback module (134) providing corrective feedback to the user through; quality assessment result quality If it meets the threshold, the PPG time series preprocessing module (140) and / or machine learning Transfer to the classification module (160); calibration / confidence score of the classification output Generating a calibrated confidence score with module (162); and classification result, quality information, confidence score, 15 heart rate, symptom reporting (230) or at least two of the following suspicious recording parameters reporting, alarm through hybrid rule / verification and alarm engine (170) which evaluates together, at least one of the following steps: clinician panel (180) review or ECG verification referral A method characterized by including the initiation steps.

2. The method according to claim 1 is the signal quality evaluation gate (130), finger contact, light 20 saturation, dropped square rate, signal amplitude, signal-to-noise ratio, pulse interval coherence, signal shape weighted at least two of the following motion metrics: stability and accelerometer / motion sensor (116) based metrics, Quality index calculator (132) through a combination of rule-based or machine learning-based quality A method characterized by index calculation.

3. Method according to claim 1 or 2, where the accelerometer / motion sensor (116) data in the PPG signal 25 not to correct motion artifacts, but to make a decision to accept or reject the measurement. The method being characterized.

4. Method according to any of the previous requirements, signal quality evaluation gate (130) It can make quality acceptance / rejection decisions on mobile electronic devices (110) without requiring cloud connection. 30 characterized by its operation as a rule engine or quantized low-resource consumption model. the method used.

5. The method is based on any of the previous requirements, and the rhythm is used when the quality threshold is not met. no classification result is produced; instead, the user feedback module (134) is used to provide feedback. holding it steady, adjusting finger pressure, changing lighting conditions, camera contact 35 by providing the user with at least one of the instructions for correction or restarting the measurement. The method being characterized.

6. This method is based on any of the previous requirements, and the number of quality rejections is predetermined. If the number is reached, the measurement session is terminated and a timestamp of the rejection reason is recorded on the device. reporting, database and audit log module (190) with information or quality criteria A method characterized by the inclusion of the step of storing the information as a traceable record. 40 7. The method is according to any of the previous requests, and the preprocessing module (140) is re-processed. sampling, polyphase resampling, basal shift removal, high-pass filtration, low band-pass filtering, Butterworth filtering, Hampel filter, outlier. reduction, windowing, Z-score normalization, smoothing, or pulse peak identification. A method characterized by including at least one of the following processes. 45 8. Method according to any of the previous requests, feature extraction module (150) pNN50, SDNN, Poincare SD1, Poincare SD2, pulse interval variability, pulse wave amplitude variability, waveform form morphology consistency, ascent / descent slope, pattern irregularity, or intra-measurement quality A method characterized by producing at least one of the following statistics. 9 9. This method is based on any of the previous requirements and is a machine learning classification module. (160) possibility of atrial fibrillation, possibility of normal sinus rhythm, possibility of other rhythm irregularities or by producing a rhythm classification result that includes at least one noisy / uninterpretable measurement label The method being characterized. 5 10. This method is based on any of the previous requirements and is used by the machine learning classification module. (160) one-dimensional convolutional neural network, 1D-ResNet, temporal convolutional network, residual connected neural network, attention-based layer, decision tree, support vector machine, or a hybrid of these A method characterized by being performed with at least one of the following combinations.

11. This method is based on any of the previous requirements and uses a PPG-based machine learning model. Pre-training with source data sets and fine-tuning or clinical with PPG-ECG recordings. A method characterized by being configured in a way that allows for adaptation.

12. The method is according to claim 11, and involves layer freezing, partial freezing during fine-tuning or clinical adaptation. Layer freezing, low learning rate retraining, temporal convolutional cap reconstruction training or at least one of the Elastic Weight Consolidation-based adjustment steps is 15 a method characterized by its application.

13. The method is based on any of the previous requirements, and the calibration / reliability of the raw model output is determined accordingly. score module (162) by Platt scaling, isotonic regression, temperature scaling or equivalent A method characterized by converting the probability calibration method into a calibrated confidence score.

14. The method is in accordance with claim 13, and calibration is repeated after each model update. execution; if the calibration success criteria are not met, the relevant model version will be used in clinical decision support work. A method characterized by not being included in the flow or by stopping the distribution line.

15. Method according to claim 13 or 14, calibration result, validation report, data set version and model version is associated with the model lifecycle / MLOps module (210) A method characterized by its concealment. 25 16. The method is based on any of the previous requirements, and is a hybrid rule / verification and alarm engine. (170) not using the classification result alone; heart rate threshold, quality index, confidence score, Symptom reporting (230), number of suspected results in the last twenty-four hours, time of measurement, pNN50, At least one of the SDNN or Poincare SD1 / SD2 parameters must be included with the classification result. A method characterized by its evaluation. 30 17. Method according to claim 16, used in hybrid rule / verification and alarm engine (170) thresholds for clinical validation, age, population, device type, device matrix, or performance monitoring. A method characterized by its ability to be determined or updated based on its results.

18. Method according to claim 16 or 17, with a predefined low or high heart rate threshold. exceeding, resulting in more than one suspected rhythm in a specific time interval, symptom reporting (230) 35 If an accompanying low confidence interval occurs or PPG-ECG discrepancy is detected, the patient limited alert to interface (182), notification to clinician panel (180) or PPG-ECG time matching The method is characterized by initiating ECG verification guidance via module (200).

19. The method according to any of the previous requests, with the T1-T2 time interval of the PPG segment. PPG-ECG 40 using the T1±Δ and / or T2±Δ tolerance range of the ECG reference recording / segment (202). matching with time matching module (200) and validation tag from matching segment, training candidate label, conflict information or confidence score review information (204) generation step a method characterized by its inclusion.

20. The method according to claim 19 is PPG-based classification result and ECG reference record / segment. (202) If there is a discrepancy between the reference label or the confidence score is 45 below the predetermined threshold If the value is low, the relevant record is referred to the clinician panel (180) for review. The method being characterized.

21. Method according to any of the previous requirements, asynchronous explainability module (220) calculates the explainability output asynchronously after the main classification response and the patient without presenting an explainability map to the interface (182) that would give the impression of a definitive diagnosis to the clinician panel (180) A method characterized by demonstrating for the purpose of retrospective review or decision support. 5 22. The method according to claim 21 is the explainability output of the asynchronous explainability module (220). SHAP should generate it using TCN Attention Rollout or an equivalent significance / explainability method, and Characterized by transmitting to the clinician panel (180) via WebSocket or an equivalent communication channel method.

23. Method according to any of the previous requests, and 10 that did not trigger an alarm or clinical investigation. The raw PPG record is retrieved from local memory / temporary storage (118) after a predetermined period of time. automatic deletion; occurrence of alarm, low confidence, discrepancy or need for clinical review. This method is characterized by storing the relevant record in an encrypted database.

24. Method according to any of the previous requests, raw data, processed data, quality rejection reason, Reporting, database and audit log 15 of alarm log, clinical validation data and training candidate data. The module (190) is characterized by being subject to different storage, deletion or encryption rules. the method used.

25. Method according to any of the previous requests, including measurement initiation, measurement refusal, reason for refusal, Classification result, confidence score, alarm generation, clinician imaging, clinical note addition, ECG At least one of the following operations must be completed within 20 days: matching, model update, data export, or data deletion. by recording as audit log through reporting, database and audit log module (190) The method being characterized.

26. A method according to any of the previous requests, and a classification made by the clinician. Correction information: clinician ID, timestamp, previous classification, corrected classification, model 25 characterized by being stored in association with the version, dataset version, or validation record. method.

27. The method according to claim 26 is that clinician corrections reach a certain number or a certain number. If time elapses, the model update candidate is determined by the model lifecycle / MLOps module (210). initiating the workflow; that workflow includes data retrieval, preprocessing, model training or fine-tuning, calibration, validation, bias control, shadow mode comparison, human approval, limited distribution or 30 A method characterized by including at least two drift / performance monitoring stages.

28. Method according to claim 27, calibration or validation in the model update candidate workflow. If the criteria are not met, the relevant model version will not be released or the previous model will be used. A method characterized by the continued use of the current version.

29. An arrhythmia configured to perform the method according to any of claims 1 to 28. 35 decision support system (100); mobile electronic device containing camera (112) and light source (114) (110); PPG signal acquisition module (120); signal quality assessment working on mobile device gate (130); user feedback module (134); preprocessing module (140) or classification Preparation module; Machine learning classification module (160); Calibration / confidence score module (162); and with mandatory core modules including hybrid rule / verification and alarm engine (170), 40 Clinician panel (180), reporting, database and audit log module (190), PPG-ECG time matching module (200), asynchronous explainability module (220) or model lifecycle / MLOps The system is characterized by including at least one optional support module from module (210).

30. According to claim 29, the system is (100) and the signal quality operating on the mobile electronic device (110). Low quality measurement of the evaluation gate (130) is sent to the cloud or master rhythm classifier 45 rejecting it before sending and quality measurement to the preprocessing module (140) or machine learning The system is characterized by its transfer to the classification module (160).

31. System (100) according to claim 29 or 30, clinician panel (180) patient-based final measurement time, rhythm classification, heart rate, quality index, confidence score, symptom reporting (230), PPG 11 waveform, ECG reference record / segment (202), clinical note, explainability output or alarm A system characterized by displaying at least one image from its past.

32. System (100) according to any of claims 29 to 31, where patient, clinician, institutional manager or Role-based access control, session timeout, or 5 that includes at least two of the system administrator roles. A system characterized by its encrypted storage mechanism.

33. System (100) according to any of claims 29 to 32, model lifecycle / MLOps module (210) dataset version, source code version, model weights, hyperparameter information, Calibration result, validation result and distribution version hash or equivalent cannot be changed. A system characterized by its association with records. 10 34. When executed by one or more processors, request 1 through 28 commands that execute the specified quality-gated PPG-based arrhythmia decision support method steps a non-volatile, computer-readable recording medium containing.

Images