Patent Application
20220230758
The invention relates to the detection of the risk for a patient for having a torsade de pointes event, and the mechanisms underlying such risk, in particular via the use of neural networks.
Torsades de Pointes (TdP) is an arrhythmia that can deteriorate to ventricular fibrillation and sudden death. It occurs essentially in the congenital long QT syndromes or as a rare side effect of QT-prolonging cardiac and non-cardiac drugs.
The current detection strategy is based on the QT measurements (i.e., interval between the beginning of the QRS complex and the end of the T wave, traditionally, using generally the lead II of the electrocardiogram (ECG)). TdP risk is a major issue for the drug industry, and can be the reason for withdrawing drugs from market.
Another method was described in US 20190059764, which discloses an algorithm using the QT interval corrected with respect to the heart rate, the maximum amplitude of the first peak of the T wave and the interval between the maximum first peak of the T wave and the end of the T wave. Although efficient, this algorithm may not always be easy to implement, as it requires comparing the ECG before and after intake of a QT prolonging drug.
It is important to find novel ways to quantify the risk of developing TdP from drug intake. Indeed, looking at the QT only is not sufficient to detect intake of a drug which prolongs QT as is the case for Sotalol. Sotalol is a potassium-current inhibitor drug, mainly used to prevent atrial fibrillation, angina pectoris and hypertension. Sotalol can lead to a prolongation of the ventricular repolarization and, rarely (1.8-4.8%), to TdP.
Recent advances in the fields of machine learning (ML) and more specifically Deep Learning (DL) have demonstrated ability to process with bio-signals including electrocardiograms. Specifically, deep learning has been used for several applications using ECG signals. In particular, convolutional neural networks (CNN) architectures were applied to detect various types of arrhythmia (Acharya, Oh, et al. 2017), myocardial infarction ((Acharya, Fujita, Lih, Hagiwara, et al. 2017); (Acharya, Fujita, Oh, et al. 2017); Liu et al. 2018), coronary artery disease and shockable ventricular arrhythmia. Most of these studies used a single-lead approach and no clinical data were included in the models.
Another recent study trained a model on single lead ECG end-to-end to detect up to 12 different arrhythmias. The database included recordings from more than 29000 subjects annotated by certified cardiographic technicians. The model tested with 336 records from 328 patients outperformed the average score of a group of 6 board-certified cardiologists (Rajpurkar et al. 2017). The same team repeated the analysis with a dataset of more than 53.000 subjects. The model tested on an independent dataset obtained an area under the curve of 97%, whereas the average cardiologist got 78% (Hannun et al. 2019). End-to-end models were also used by Pourbabaee, Roshtkhari, and Khorasani 2017, who studied paroxysmal atrial fibrillation.
Other studies focused on models with multi-lead ECGs input. For the detection of generalized anterior myocardial infarction, Liu et al. 2018 used a 4-lead approach that led to accuracies >90% with a 5-fold cross-validation. Tison et al. 2018 created instead a Hidden Markov Models-CNN model that took as an input 12 leads in order to detect pulmonary arterial hypertension, hypertrophic cardiomyopathy, cardiac amyloid and mitral valve prolapse. The area under the ROC curve was in the 77%-94% range for the 4 classes. 12 leads were also used by Attia et al. 2019, who applied a CNN model on a database of >97.000 patients to detect left ventricular dysfunction. They used a holdout dataset of >52.000 patients and an overall accuracy of 86%. Moreover, a subset of the patients that were erroneously classified as ventricular dysfunction, developed in the years to come a low ejection fraction, suggesting that the model was able to detect this condition before it becomes visible.
Attia et al., 2018 tried to estimate plasma drug concentration using CNN by analyzing 10-s recordings of 42 patients that received Dofetilide (antiarrhythmic drug) or placebo. These recordings were recorded at specific time points before and after drug consumption. In this experiment the authors used the data from two distinct prospective randomized controlled trials available in the Physionet repository. In their regression analysis, using CNN, they reached a correlation of 0.85. However, the database was relatively small (42 subjects) and no cross-validation in training was used, leading to possible overfitting. Furthermore, this document doesn't disclose that such system is actually able to detect that a patient will soon have an episode of torsade de pointes, as the reported data only focuses on the increased QT, which is one factor for TdP but not sufficient to conclude on the actual real risk of occurrence. This article further used the ECG data from 12 leads, whereas the methods herein disclosed show that the methods can be operative with ECG generated with fewer leads (minimum of one). Furthermore, the authors didn't demonstrate that it is possible to differentiate the kind of congenital long QT. Finally, the authors mention the limitations of the study, and in particular that no conclusions can be drawn with regards to the occurrence of arrhythmic risk.
Yildinm et al. 2018 were able to detect 17 types of arrhythmias with an accuracy of 91% (test set) using an end-to-end model. However, this analysis was based on 1000 single-lead ECGs of 10 seconds recorded from 45 individuals. This means that the same subject had data both in training and testing, leading to non-independent datasets and overfitted models. Other authors applied data augmentation in order to increase the dataset to detect 5 heart conditions and accuracy was above 90% as tested with 10-fold cross-validation procedure (Acharya, Oh, et al. 2017). These authors also compared the performances between a denoised and raw database, showing that noise removal was not associated to a higher score. However, the cross-validation random split was done after the data augmentation. The fact that the signals were not completely independent between the training and testing set might lead to overfitted performances. These authors don't report either that they are able to discriminate the nature of congenital long QT or predict the risk of occurrence of TdP in near future, such as within 48 hours or in the following 24 hours, as shown in the present disclosure.
It is important to find novel ways to quantify the risk of developing TdP from drug intake. Indeed, looking at the QT only is not sufficient to detect intake of a drug which prolongs QT, such as sotalol, as shown in the examples.
Due to the rarity of the TdP event, and to the fact that this event occurs essentially after intake of QT-prolonging drug, the strategy developed by the inventors was to develop a model that is able to detect the intake of sotalol (which is a marker of the risk to have a torsade de pointes). Using the model so developed, the inventors were able to show that it is able to discriminate between congenital long QT syndromes (an inherited disorder of cardiac repolarization that predisposes to syncope and to sudden death from polymorphic ventricular tachycardia, i.e TdP), and in particular to identify LQT2 patients vs. LQT1 or LQT3 patients (for which the biological mechanisms at hand are different).
It is to be noted that the risk mentioned above is not an absolute risk, but rather a relative risk. The patient detected “at risk” in the context of the present invention presents a higher risk than the population as a whole. In other words, the patient presents characteristics that are shared by other patients who have recorded a TdP episode. This means that the risk is linked to the sensitivity of detection of the fact that the patient will have TdP. However, since the test is not 100% specific, some patients may present the characteristics and not have a TdP episode (false positive) and since the test is not 100% sensitive, there may be some patients who don't have the characteristics and that will develop a TdP event (false negative). The quality of the test is represented by the Area Under Received Operating Curve (AUROC).
The methods and systems herein disclosed are able to classify the patient in classes (increased risk/no increased risk) and in subclasses pertaining to the type of risk (or mechanism by which the risk is present) for patients in the increased risk class (drug induced risk, acquired risk, congenital long QT syndrome and type thereof).
The invention relates to a method for producing a machine learning algorithm (or machine-learning system) capable of predicting or estimating the risk for a patient to have a torsade de pointes event, and preferably the cause (or underlying mechanism) thereof, comprising:
As demonstrated in the examples, the inventors have shown that using electrocardiograms from patients make it able to classify the patients as having a risk for torsade de pointes (as correlated with the intake of sotalol), and is able to classify the mechanism underlying the risk (intake of a drug, congenital long QT syndrome, type of congenital long QT syndrome). The fact that it is possible to obtain a system that makes it possible to detect such risk (and the mechanism underlying the risk), using the data set as disclosed above was surprising for the inventors. Indeed, the inventors show that the ECG signal contains information that can be extracted from the machine-learning system and allows a very detailed classification of the patients for the TdP risk.
In particular, the inventors have shown that using such system makes it possible to detect whether the patient will have an episode of torsade de pointes in a short delay (within 48 hours or 24 hours). The inventors also showed that the system can also be operated using ECG obtained from a single lead.
As such, the output of the system that is trained may rely on a classification technique, which provides a prediction, probability or likelihood, rather than certainty, that the patient has a certain condition (here to risk a TdP event, and the nature of the biological mechanism). For instance, classifiers using one or more of the following techniques may be used to determine the likelihood of such risk: linear discriminant analysis, neural network, clustering techniques, support vector machine, logistic regression, naive Bayes, random forest, decision tree, etc.
It is, however, preferred when the machine learning classifier is an artificial neural network, and preferably a convolutional neural network.
Thus, and more specifically, the invention relates to a method for producing an artificial neural network system capable of predicting or estimating the risk for a patient to have a torsade de pointes event, and preferably the cause (or underlying mechanism) thereof, comprising:
Using such input data and training the machine-learning system, preferably the neural network, allows to obtain models that are able to determine whether an ECG of a patient presents the same characteristics than the ECG of patients at risk of having a TdP event. Such models can thus be used for implementing various methods as disclosed below.
The invention thus relates to a computer-implemented electrocardiogram (ECG) analysis method comprising:
This method, as well as other methods herein disclosed is performed ex vivo.
The invention also relates to a computer-implemented electrocardiogram (ECG) analysis method comprising:
In particular, the machine-learning model (or system) is a neural network, such as a feedforward neural network, a convolutional neural network, or a recurrent neural network, this list being non limitative. It is preferably a convolutional neural network.
Step a. corresponds to the extraction of a segment (part) from an ECG signal (preferably during a 10-second window) in order to obtain a data segment that can be used for processing in a machine-learning system that has previously been trained as disclosed above. It can be done online or offline.
In step b., it is possible to normalize and/or to standardize the data segment or no processing at all. For instance, when the ECG signal has been acquired on multiple leads, the standardization can be performed using the mean of the signals acquired for each lead for the patient or per each lead. Normalization may be performed using the ECG signals of all patients that are present in a reference database or for each ECG.
In step c., the data segment is applied to the machine-learning algorithm that has been developed as disclosed above, in order to obtain an output. Other data can also be applied together with the data segment (such as sex, age of the patient, kaliemia, etc. . . . ). When the data has been obtained on multiple leads, the data from all the leads can be entered at once as the ECG signal (composite ECG signal) or part of or all leads separately.
The invention also relates to a method for estimating the risk for a patient to have a torsade de pointes event and underlying mechanism thereof, comprising:
For the methods disclosed above (and throughout the present disclosure), the output provides a likelihood of the patient belonging to the classes that were defined during training.
As illustrations
The risk of having a torsade de pointes event is linked to the belonging, for the patient, to a class in which patients are more at risk. In other words, the risk is a relative risk as compared to patients classified as control. The risk of having a TdP event is considered to be increased (as compared to control patients) when the ECG of the patient is classified like an ECG of a patient belonging to a class of patient with increased risk (patient taking a drug that increases the QT, or patient with a long QT syndrome). The machine learning algorithm is also able to discriminate the class and thus to provide information as to the underlying mechanism that increases the TdP risk: inhibition of IKr channel for patients taking sotalol or LQT-2 patients, mutation in the KvLQT1 gene (also known as KCNQ1) involved in the IKs channel for LQT-1 patients, mutation in SCNSA, the gene encoding the sodium channel INa/INa late for LQT-3 patients.
It is thus possible to act adequately for the physician (stop medication and check that ECG becomes normal again, request genetic confirmation for long QT syndrome patients and prescribe adequate medication).
The invention also relates to a cardiac monitoring device comprising:
The standardization/normalization of the ECG signal in ii. is optional and can be performed as disclosed above. Step iii. Aims at using only the signal for a specific duration (generally 10 seconds) in the machine learning classifier. This step is optional and is not always necessary.
As indicated above, the machine learning classifier is preferably a neural network, in particular a convolutional neural network.
The invention also pertains to a system for the evaluation of the risk for a patient to have a torsade de pointe event, comprising:
It is reminded that, when multiple leads are used, the ECG data is the vector formed by all these multiple ECG signals from the different leads. The examples provide an illustration of such vector for 8 leads.
In the above embodiments, the defibrillator or monitor comprising the at least one electrode and the computer/processor or processing unit can be at different locations. In this embodiment, the ECG signal is sent to the processing unit or processor in order for the machine learning classifier to process it. It is also possible to send the output to the defibrillator or monitor or to another device (such as a smartphone). In this case, it is possible to perform a monitoring at home of the risk of having a torsade de pointe event, or its evolution (depending on the evolution of the likelihood calculated by the machine learning classifier). When the defibrillator or device having the at least electrode and the processing system are distant, it is preferred when communications between these systems are encrypted.
The invention also makes it possible to evaluate the torsadogenic potential of a drug, of a product or of a composition of interest.
It is indeed important to be able to evaluate the torsadogenic potential of a compound, during clinical trials prior to the obtaining of a marketing authorization, in order to be able to draft the possible precautions for use of this drug. The torsadogenic potential of a substance is the potential of this substance to induce a torsade de pointes after administration in a patient.
The FDA (Food and Drug Administration) and the EMA (European Medicines Agency) have published guidelines resulting from the International Conference for Harmonization (ICH) on how to carry out these trials, under the title E14 Clinical Evaluation of QT/QTc Interval Prolongation and Proarrhythmic Potential for Non-Antiarrhythmic Drugs. These guidelines are accessible on the website of this agency, and are accompanied by questions and answers.
The url address of the guidelines is:
http://www.fda.gov/downloads/drugs/guidancecomplianceregulatoryinformation/guidances/ucm073153.pdf
As discussed above, the QT interval is an imperfect biomarker for the risk of proarrhythmic effect. In particular, the examples show that the AUC obtained when using QT for detecting patient having had sotalol is 0.76, which proves that improvement can be obtained.
It is therefore required to rigorously characterize the effects of new drugs with regard to the risk of delaying cardiac repolarization. The machine learning classifier defined above makes it possible to detect whether such risk is present, by analyzing the ECG of patients having been administered the compound of interest.
The invention thus relates to an ex vivo method for determining the risk for a substance to induce a torsadogenic effect after administration to a patient, comprising the following steps:
It is preferred when multiple ECG signals, acquired at various time points after administration of the substance of interest, are applied to the machine learning classifier.
It is also possible to apply, to the machine learning classifier, ECG signals for the patient, acquired before administration of the substance of interest.
Generally, and in order to limit patient variability, the method disclosed above is repeated on a cohort of patients containing a number of patients greater than or equal to 2. Carrying out these studies on a sufficient number of patients thus makes it possible to obtain results which present a statistical reality for the molecule that it is desired to test (drug of interest), eliminating the inter-patient variations.
A placebo (negative control), a positive control (such as sotalol at 80 mg) and the substance of interest can be compared.
Sotalol at 80 mg is indeed preferably used as positive control. This molecule, at this concentration, has many advantages:
When testing the substance on females, it is advisable to carry out the various measurements at the different time of the menstrual cycle. This is because it is known that the level of hormones influences the risk, in a woman, of repolarization disorders (long QT).
In general, as indicated above, the above steps are carried out for each patient of a cohort of patients, and the risk is thus obtained for each patient, and for each substance.
It is preferable for the cohort of patients (the number of patients on which the substance of interest is tested) to be at least 10 patients, preferably at least 20 or even 50 or 100 patients. Those skilled in the art will be capable of determining the minimum number of subjects in the cohort of patients in order to obtain results which are statistically significant.
Likewise, before carrying out these studies, it is preferable for those skilled in the art to have carried out pharmacokinetic studies and dose-response studies on the substance of interest, in order to determine the best times to perform the ECG acquisition and machine learning classification.
The invention also relates to a method for classifying (determining) the nature of a congenital Long QT syndrome in a patient, comprising the steps of
In particular, this method can be performed on a patient, which has been detected as having a congenital long QT syndrome, but for which the genetic analysis has not yet been performed.
This method thus makes it possible to adapt a treatment very early before the genetic confirmation is obtained (as it can take a few months).
It is reminded that approximately 75% of patients with a clinically certain LQTS (Long QT Syndrome) diagnosis have mutations in one of three major LQTS—susceptibility genes (KCNQ1, KCNH2 or SCN5A) with mutations in KCNQ1 being responsible for about 35% of LQTS (type 1, LQT1), KCNH2 mutations being approximately 30% LQTS (LQT2) and SCN5A mutations being about 10% of LQTS (LQT3).
Other genes (AKAP9, CACNA1C, CALM1, CALM2, CAV3, KCNE1, KCNE2, KCNJ5, SCN4B, SNTA1) are minor LQTS genes.
For a patient for whom a genetic test has been performed and for whom mutations have been found in some of these genes, the machine-learning classifier system will be able to detect whether or not such patient has actually been classified as LQTS by analyzing the ECG of such patient. The system herein disclosed thus makes it possible to identify whether a given mutation classified as a variant of unknown significance is more likely to be pathogenic or benign.
The invention thus relates to a method for determining whether a mutation of interest in a gene is susceptible to be associated with a congenital LQTS, comprising
It is reminded that the output of the machine-learning classifier is a likelihood that the patient has prolonged QT (when the machine-learning classifier has been trained with ECGs from patients having received QT-prolonging drugs and from patients not having received QT-prolonging drugs), and that it is thus possible to classify the patients in one of these two classes; if the patient is classified in the class of patients having received QT-prolonging drugs, it is considered to have a prolonged QT. If the patient has a prolonged QT, if the mutation is thus susceptible to be associated, linked or causative to this condition. If the patient has not a prolonged QT, the mutation is likely not to be causative or associated to LQTS. As disclosed above, the result will be more specific if the training of the neural network was performed with classes that are specific of congenital Long QT and of their type 1, 2, 3 or other). However, as shown in the examples, the training with only data from patients with QT-prolonging drugs also made it possible to identify congenital long QT.
In particular, the mutation of interest is in a gene as cited above (either one of the main three genes, or in a minor gene).
It is also reminded that such testing must be understood as probabilistic rather than unconditionally deterministic and that future work (such as QT analysis for other patients with the same mutation) are needed to confirm the role or association of the mutation to LQTS. However, being able to distinguish harmful from neutral variants would help avoid unnecessary interventions in carriers of benign alleles, and focus on patients with truly relevant mutations.
Since the disclosure also shows that the methods disclosed herein are able to identify the nature of LQTS, it is also possible to treat the patients adequately when the nature of the LQTS is identified by the machine-learning classifier. For example, beta-blockers (class I or Class IIa) such as Atenolol, Nadolol, Propranolol can be used for LQTS (especially for LQT1 and LQT2), while Mexiletine, Ranolazine or flecainide (non-selective voltage-gated sodium channel blockers) can be used for LQT3
The invention also relates to a method for detecting (or estimating) the risk for a given patient to present an episode of torsade de pointes, comprising the steps of:
In a specific embodiment, the ECG data is sent to a remote machine learning classifier by electronic/network means (such as cabled or wireless network, GSM, 3G or wifi) and the output is sent to the patient and/or a physician device (phone, computer) by electronic/network means (such as by email or by a Short Message on a phone device (SMS)). In particular, it is sent to the physician so that the physician can thus determine the risk of TdP for the patient and act accordingly. In particular, such method is performed daily. It is preferred when communications between the machine measuring the ECG and the remote server containing the machine learning classifier is encrypted and between the remote server and the patient or physician device are encrypted.
It is also possible to classify or assign the patient in a class (nature of the long QT syndrome, increased risk of TdP event) by using a voting system as disclosed below. As shown in the examples, such system will increase the accuracy of the methods.
In a preferred embodiment, the likelihoods (probabilities or estimates) for a patient are calculated by averaging risk scores obtained when multiple ECG signals for the patient are applied to the machine learning classifier (voting system). The repetition of the machine learning classifier on multiple ECG samples is equivalent to a physician looking at different ECG samples to reach a conclusion. In fact, this method makes it possible to classify the patient in a given class, which is the one that is the most probable in view of the different likelihoods calculated by the machine learning classifier. It is possible to process multiple ECG from the same patient were processed by the models and to affect the patient in a class when the majority of the outputs indicate a higher probability of the patient of being in that class.
It is also possible to classify the patient in a given class by using different machine classifiers that have been independently trained for the same purpose. Using this kind of voting system is equivalent of multiple physicians looking at the same data before reaching a decision.
It is also possible to combine these two voting approaches.
It is reminded that the machine-learning system makes it possible to provide a probability of the patient being in the classes that have been used during the training process. During the operational phases, it is important to be able to assign a class to a patient (or a patient to a class).
The methods above can be completed by a method, wherein the patient is assigned to a class by
It is also possible to use a method, wherein the patient is assigned to a class by
The ECG signal can be acquired on a single lead (preferably the V3 or DII lead). It is however preferred when the ECG signal is acquired from more than one lead, preferably from at least 3, more preferably from at least 6 or from at least 8 leads. Any lead can also be used individually. DI, V4 or V6 leads are of particular interest.
The inventors have determined in particular that using a machine-learning classifier, that has been configured (taught) with ECG coming from patients having been administered a QT-prolonging drug and with patients to whom no QT-prolonging drug has been administered, makes it possible to actually detect whether a torsade de pointe episode is to occur within 48 or 24 hours, or to detect whether a torsade de pointe episode will not occur within 48 or 24 hours.
It is important to understand that the risk of having a torsade de pointes episode is not strictly correlated with a prolonged (or enhanced, or increased) QT. Although such prolonged QT increases the risk (in the sense that there is an increased relative risk in the population of patients with increased QT vs. a population of patient with normal QT), such increase by itself is not sufficiently predictive of the fact that a torsade de pointes will actually occur.
The population risk (increased risk in the population) doesn't provide by itself the information for a given patient in the population.
However, it is important for the physician to be able to assess whether an individual patient is at risk or not of having a torsade de pointes episode in the near future. It can also be said that it is important to determine whether the individual patient will not have any torsade de pointes episode in the near future. The patient is generally a patient to whom is administered a drug that is known to increase the QT and hence to increase the risk of torsade de pointes events. The duty of the physician is to monitor the patient in order to decrease its individual risk. Monitoring may be by a constant hospital surveillance which is costly and occupies beds in the hospital. It would thus be preferable for the physician to be able to rule out the fact that the patient will indeed have a torsade de pointes episode in the near future (within 48 or 24 hours) and to only closely monitor the patients that are the most at risk.
The inventors have shown that the training of the machine-learning system with a database as disclosed in the examples makes it possible to obtain an output that goes beyond the training, when new ECGs are provided, and that is able to determine a risk of torsades de pointes with high specificity.
It is reminded that
(1) sensitivity is the probability that the diagnosis is positive in individuals having the condition (detection of true positives): the test is positive if the patient is having the condition.
(2) specificity is the probability that the diagnosis is negative in the individuals not having the condition (there is non-detection of true negatives): the test is negative if the patient is not having the condition.
(3) Positive predictive value (PPV): is the probability of having the disease if the diagnostic test is positive (i.e. that the patient is not a false positive): the patient is having the condition if the test is positive.
(4) Negative predictive value (NPV): is the probability of not having the condition if the diagnostic test is negative (that the patient is not a false negative): the patient is not having the condition if the test is negative.
In the present case, the prevalence of the condition (actual torsades de pointes event) is low. However, with a high sensitivity, the NPV will be high. It is thus possible to say that a person has not any risk (ir a very little risk) of having a torsade de pointes event in the near future (within 48 or 24 hours) if it is not detected as such by the system. In other words, the inventors showed that the system is very good at detecting patients who will not have a torsade de pointes due to the high sensitivity, even though some patients may be detected as false positive due to a specificity that is not as high as the sensitivity. It is thus possible to reduce the number of patients to monitor, and focus on a subgroup for which there is a higher prevalence of TdP event (higher risk).
Furthermore, the inventors showed that the test is semi-quantitative (i.e. the higher the score, the higher the risk). For patients above a given threshold, the individual risk will increase and one can increase the monitoring, stop the administration of the drug that is at risk (or decrease the dosage), and/or correct any possible hydroelectrolytic troubles by balancing the magnesium or potassium blood concentrations, maintaining or getting the patient in the hospital for surveillance.
As disclosed below, the machine-learning classifier (preferably a neural network) provides, as the output, a score. The inventors have shown that the higher the score, the higher the risk of a torsade de pointes event within 48 or 24 hours, and that it is highly specific. It is to be noted that the value of the score doesn't fully correlate with the increase in the QT: hence, the classifier is able to classify the patients using other elements than merely the QT length. The information provided by the model (as the score) is thus different from the information delivered by the measure of the QT length.
It is possible to define three classes, using two cutoffs. Below one cutoff, the patient's individual risk will be very low, and the monitoring can thus be reduced.
For patients above another threshold, the monitoring will be intensified. For patients between the two thresholds, the monitoring will be maintained as usual.
The invention also relates to a method for determining whether a subject is at risk (or at increased risk) of having an episode of torsade de pointes within 48 hours, comprising
The invention also relates to a method for determining whether a subject is at risk (or at increased risk) of having an episode of torsade de pointes within 24 hours, comprising
It is to be noted that the inventors have trained the model so that the risk of torsades de pointes is higher when the score is higher. This is completely linked to the annotation of the data as provided to the machine learning classifier at the time of training. Using a different annotation could lead to another model in which the risk would be higher when the score is lower. The point is that the comparison is to be made with a threshold, which enables the classification of the patient as “at risk” or “not at risk” and that the threshold and the classification will depend on choices made for the specificity, sensitivity, and type of training of the system.
In summary, the present disclosure shows that
The training database included 9014 ECG signals from 792 healthy patients, before (n=4014; Sot−) and after (n=5000) Sotalol intake (80 mg per os, Sot+). The ECGs were taken at different time points (1, 2, 3 and 5 hours after sotalol intake).
A subset of the dataset was isolated to test the trained models composed of 198 healthy patients with ECGs before (n=999) and after (n=1238) sotalol intake.
Moreover, a third dataset composed of long congenital QT patients was used to test the trained models, specifically composed of LQT-1 (n=266 patients and n=560 ECGs), LQT-2 (n=188 patients and n=456 ECGs) and LQT-3 (n=33 patients and n=67 ECGs).
Cohort Description
Two cohorts of patients were used.
The first cohort contained electrocardiograms from 990 healthy subjects before and after the consumption of 80 mg of Sotalol at different periods of time (namely at 2, 3, 4 and 5 hours after Sotalol intake). Each patient had a different number of recordings, as these recordings were either repeated or not at each time interval. Each subject had at least one recording at basal and T3. Sotalol concentration was measured in serum, at 2, 3 and 5 hours after intake, but this was not available for all subjects. Before the Sotalol intake, all signals were associated to a zero concentration of Sotalol. This first cohort was used to train the CNN (Convolutional Neural Network) and ResNet models that classify individuals in Sot+/Sot− groups and predict the concentration of Sotalol in the blood.
A second cohort composed of long-QT congenital patients was also used, specifically composed of LQT-1 (n=266 patients and n=560 ECGs), LQT-2 (n=188 patients and n=456 ECGs) and LQT-3 (n=33 patients and n=67 ECGs), from a second study to test the trained models.
Convolutional Neural Network Architecture
Two types of models were developed, depending on the inputs. Either just the ECGs (M1) or the ECGs with the data associated to the sex, age and Kalemia (M2) were used. The models' architecture is the same.
The models were based on convolutional neural networks (CNN), composed of 11 blocks of convolution and each block had two Conv1D layers with the same number of filters. The number of filters for each block were 8, 8, 8, 16, 32, 64, 128, 256, 512, 1024 and 2048 respectively (11 blocks in total). Zero padding was used for both Conv1D in order to have an output with the same length of the input (option “same”). Adam optimizer, binary cross-entropy loss and early stopping (patience=50) were used. The class weights were computed to balance the output classes before training. The remaining model parameters had default values. In the M2 model, the clinical data were concatenated to the flattened layer output, before the fully connected neural network.
Different deep learning models based on the same CNN architecture were developed. The first (M1), included only ECG data before and after Sotalol intake (respectively denoted below as Control and Sot+), while the other (M2) also included basic clinical information, namely, age, sex and kalemia quantified at study baseline.
Different models M1 were developed. They all presented the same whole performance, although with some differences in assessment for some individual during the testing phase.
Preparation of the ECG Data
Raw ECG was used, acquired with the Cardioplug device (Cardionics Inc). The ECGs were associated to 8 leads (I, II, V1, V2, V3, V4, V5, V6). Each recording lasted 10 s, with a sample size of 500 Hz. The data in the format of .SCP files were parsed using the Biosig library (http://biosig.sourceforge.net/index.html).
Similarly, raw ECG data from the long-QT patient was used, acquired with the CalECG tool from 1992 to 2018. Raw .xml files were parsed using the python xmltodict library. In total, 875 files were acquired at 500 Hz, and 208 were acquired at 250 Hz, for a total of 1083 recordings. The 250 Hz signals were up-sampled to 500 Hz using a cubic interpolation (interp1d function from the scipy python library).
No filtering, nor any other transformation was applied to the data considered in these studies, except for standardization which was performed at the whole ECG level (i.e. each lead signal is standardized by the mean of all other lead signals for a given ECG). The data were stored in python dictionaries that were used for training the models.
The ECG data used corresponds to a segment of 10 consecutive seconds of a greater ECG acquisition.
For training, the data was entered as a 3 D tensor (8 leads, 5000 time points for each lead, Sot+ or Sot−). For testing the models, the data was entered as a 2D tensor (8 leads, 5000 time points for each lead).
Model Training and Evaluation
All the models explored were created using the Keras library in Python. In both classification and regression experiments, the first database was split in Training (80%) and Holdout (20%) sets. 10-times 10-fold cross-validation were performed in the Training database to estimate the generalization confidence of the algorithms. At each fold, the train subset and the test subset had distinct examples and both sets (i.e., training and testing) were normalized using the average and standard deviation of the CV training set to avoid overfitting. The final models (M1, M2 and ResNet18) were trained in the whole training set and were tested in the holdout and the external long-qt congenital cohort. Estimated CV empirical and generalization performance is also provided.
For the classification task, the training dataset included 9014 ECG signals from 792 healthy patients, before (n=4014) and after (n=5000) Sotalol intake (80 mg). A subset of the dataset was isolated to test the trained models composed of 198 healthy patients with ECGs before (n=999) and after (n=1238) sotalol intake. The classification was performed as a binary classification: no drug (Sot−) vs. drug (Sot+). The recordings were used associated to inclusion, basal and T0 as Sot− and recordings after drug intake (at T1, T2, T3, T4) as Sot+. Finally, The long-qt congenital dataset were used to test the trained models, specifically composed of LQT-1 (n=266 patients and n=560 ECGs), LQT-2 (n=188 patients and n=456 ECGs) and LQT-3 (n=33 patients and n=67 ECGs).
Voting Strategy
Performances were computed using both single-signal analysis and by averaging risk scores from multiple recordings for a given patient and condition. The output provided by the models is a probability (likelihood) of being Sot+(having ingested Sotalol).
In order to affect the patient in one of the Sot+ or Sot− classes, multiple ECG from the same patient were processed by the models and the patient was affected as being Sot+(resp Sot−) if the majority of the outputs indicated a probability of being Sot+(resp. Sot−) higher than 0.5 (50%).
Alternatively, a voting strategy can be used by processing one ECG from a patient with multiple models (trained separately, using similar architectures or different), and affecting the patient to the Sot+(resp Sot−) class if the majority of the models indicate a probability of being Sot+(resp. Sot−) higher than 0.5 (50%). Other cutoffs could also be used.
Sotalol Intake is Accurately Detected from Raw ECG Signals
The performance of the models was evaluated either at the single ECG level or averaged by patient for each of the 10 times 10-cross validation folds (CV). It was observed that the average test accuracy in CV is comparable to the holdout accuracy as is the overall empirical accuracy and the empirical in the training process (CV), indicating a good estimation of the performance of the models.
The average generalization accuracy of the model M1 is 0.885 (sd=0.003) when computed for each ECG and 0.945 (sd=0.004) when computed for each patient (using the voting strategy as disclosed above). Similarly, for the model M2, the average generalization accuracy is 0.89 (sd=0.003) when computed for each ECG and 0.947 (sd=0.003) when computed for each patient.
There is no statistical difference between M1 and M2, thus indicating that the ECG alone contains all information pertaining to the detection of the ingestion of Sotalol.
However, there was a highly significant difference between the generalization accuracy at the ECG and patient level (voting, p<1.03e-25), indicating the importance of multiple signal acquisitions.
Finally, the precision of these models is highly elevated, 0.963 (sd=0.004) and 0.961 (sd=0.003) for M1 and M2 respectively (voting) as is the recall for the same models: 0.934 (sd=0.005) and 0.937 (sd=0.004).
Altogether, these results indicate that the CNN model trained only with raw electrocardiograms is very sensitive and accurate in predicting whether a healthy subject has taken Sotalol of (80 mg) or not.
Sotalol-Intake Classification Models Discriminate Long-QT Congenital Profiles
Next, the simplest (M1) model trained to discriminate Sotalol-intake patients was used to classify long-QT congenital patients from the LQT-congenital cohort. The model provided a risk probability [0,1], which was converted in a Sot−/Sot+ class based on threshold >=0.5.
First, it was noticed that most of the Sot+ ECGs (95%) are classified as Sot+ and most of the controls (87%) are classified as Sot−. The voting results display superior performance for the Sot+(99%) and control (93%) groups as previously discussed above. Based on these observations, only the results integrating ECG signals at the patient level (voting) are later discussed.
Second, it was noticed that all LQT patients are classified in their majority as Sot+. LQT-2 display the strongest proportion (85%) of Sot+ followed by LQT-3 (67%) and LQT-1 (63%). Moreover, ROC analyses of model M1 were performed to discriminate groups together. The results for comparisons between the Sot+vs. control, LQT-1, LQT-2 and LQT-3 groups presented AUCs (Area Under Curve) of 0.99 (Confidence Intervalle, CI [0.98;1]), 0.79 (CI [0.75;0.83]), 0.59 (CI [0.53;0.64]) and 0.75 (CI [0.66;0.84]) respectively (
As a matter of comparison, the AUC between Sot+ and controls when assessing only the QT is 0.7674 [0.7381;0.7968].
The low AUC of Sot+vs. LQT-2 indicate that this model cannot discriminate well these two groups, showing their inter-similarity (IKr channel inhibition). However, the fact that the AUC was higher than 0.5 shows that there are some differences between these two conditions and that further training (with one class for Sot+ subjects and another class for LQT-2 subjects) will allow to be able to discriminate Sot+ and LQT2 patients.
Altogether with the high AUC of LQT-2 vs. control (0.92), these results indicate that the M1 model trained with Sotalol data is relevant and pertinent as a surrogate to discriminate LQT-2 patients from controls, in case of a patient with a congenital long QT.
Sotalol-Intake Classification Models Identify Patients with Torsade De Pointes at the Time of the Event
Two patients were hospitalized shortly after an episode of torsade de pointes.
ECG was performed on these patients and the ECG signals were applied to the neural network classifier as disclosed above.
For patient 1, the likelihood of being at risk of torsades de pointes (output of the model) were 0.991629, 0.989505, and 0.072291 (three ECG).
For patient 2, the likelihood were
Visit 1: 0.991812, 0.999446 and 0.984214
Visit 2: 0.998443, 0.999096, and 0.999811.
These results illustrate an example of patients who had a TdP episode and were classified as Sot+ shortly after this episode. The patients didn't have a congenital long QT and didn't take any drug inducing QT prolongation. Their QT prolongation was acquired.
The models described above are thus able to classify patients as “at risk for Torsade de Pointes”, just after a cardiac event, which makes it possible to make a retrospective diagnosis.
Discussion
The study above shows that that ECG signals carry important information and are affected by the Sotalol intake, which inhibits the IKr channel.
Indeed, being able to provide an estimation of the previous intake of Sotalol (or of another drug) in the blood may lead to a better understanding of the risk of TdP.
QT prolongation has been shown to be associated to TdP and therefore many studies considered it as a TdP surrogate. However, recent studies demonstrated that QT prolongation alone is not a reliable predictor due to the high false positive and false negative rates (Hondeghem, 2018; see also AUC for discriminating Sot+ and contrils using QT alone).
The models developed herein were able to accurately classify, with an overall generalization accuracy >94%, if the patient did or did not take Sotalol, regardless of the time after drug consumption (within 4 hours). It was shown that multiple acquisitions taken together with a voting approach improved the classification. Per se, this demonstrates presence of a rich information contained within an electrocardiogram, richer than QT only.
Next, it was shown that this surrogate model could also discriminate long-qt congenital patients at different degrees based on the type of mutation LQT-1, LQT-2 and LQT-3. This is very important as these mutations are rare and difficult to lean.
Such classification method is very useful in the case of compliance control and to adjust the dose in the context of drug monitoring. Moreover, given the importance of lowering healthcare-related costs, this system would be cheaper, more practical and faster than standard blood analysis. It would allow to also have a continuous drug-estimation instead of discrete results.
Other Developments
Various models were also developed:
When using only one lead, it was shown that the same type of results (accuracy) could be obtained than for the models described in detail above that use the ECG from 8-leads.
Another model is developed, using as the training set, ECG from controls, patients having ingested sotalol, LQT-1, LQT-2 and LQT-3 patients. This model will provide, as the output the probability for the patient to be in each of the five classes. Using a voting system (preferably by providing multiple ECG of the patient or alternatively by using multiple CNN models), it is possible to provide information for the physician and allows an appropriate care.
Models can also be developed and enriched, using multiple TdP potentially inducing drugs to be able to discriminate the drug that has been taken by the patient.
Detection of Torsades De Pointes Events
The inventors applied the model to ECGs (n=1080) from a cohort of 41 patients with followup recordings (average=26, sd=16, 1065 analyzable) and at least 1 event of TdP annotated as such by the cardiologist. The risk score was higher within the 24 hours of the TdP events versus ECG from the same individual more than 24 hours apart from the event.
This difference in risk score was significantly different (AUROC: 0.61+/−0.029,
In the model developed, the output is a score from 0 to 100. The higher the score, the higher the risk.
Table 1 provides some of the specificity and of the sensitivity values for various thresholds.
The choice of the threshold is to be made by the physician, depending on whether he wishes to have a high specificity of high sensitivity. As indicated above, it is possible to use two cutoffs or thresholds so that to have a good NPV for patients below the first cutoff, and increase the monitoring of the patients above the second cutoff.
Further training of the machine will improve the specificity without lowering the sensitivity.
This data shows that appropriate training of a machine-learning classifier (here a convolutional neural network) with data from patients having been administered a QT-prolonging drug makes it possible to detect people who will have a TdP event, or alternatively people for which one can predict that they will not have a TdP event within 24 h. This is in contrast with the conclusions of Attia et al., who say that they can't conclude with regards to the occurrence of arrhythmic risk (TdP).
Machine-Learning Classifier with One Lead
The convolutional neural network (CNN) as disclosed above was trained using the ECG data from only one lead.
Such training was made individually with each of the following leads DI, DII, V1, V2, V3, V4, V5 and V6. This thus led to obtaining 9 models (one for each individual lead and one with the composite ECG by averaging the signal from the eight leads).
Every trained system was then tested by inputting ECG data of the lead used to train it, from a testing set, as disclosed above. The classification of patients was as accurate as for the CNN trained using the ECG data from all leads (composite model) with AUROC ranging from 0.88 to 0.98.
It is also to be noted that the models were also operative when the data obtained from one lead only was applied to a model that was trained on the basis of ECG data from another lead (or for all leads). In particular, when the data obtained from the DII lead is applied to a model trained with ECG data from the DII lead, the AUROC (representing the ability to discriminate patients having been administered a QT-prolonging drug from patients not having been administered such kind of drug) is 0.98. When such DII ECG data is applied to a composite-trained model (trained with the data from 8 leads), the AUROC is 0.94. When such DII ECG data is applied to a V4 ECG-trained model (trained with the data from the V4 lead), the AUROC is 0.90. When a V4 ECG data is applied to a model trained with the data from the V5 lead, the AUROC 0.90.
This data shows that a model trained with either an ECG from a single lead or with a composite ECG obtained from multiple leads can be used with any ECG data, whether from an individual lead or composite, with an AUROC mean (taking into account all combinations of the data from composite ECG or ECG from individual leads dl; dll, V1, V2, V3, V4, V5 or V6, as the input or in the trained model) is 0.877 with standard deviation of 0.067.
These results show that the machine-learning classifier can provide good analysis output when properly trained using only ECG data from only a single lead. This opens the way to application where the patient can himself place a single electrode to measure the ECG, which can be sent to a centralized server and analyzed by the machine classifier, so as to detect the risk for the patient to have a TdP and allow the physician to act if such risk is detected.
Discrimination of Congenital Long QT (LQT)
The models trained with ECG data from a unique lead were also tested to determine their ability to discriminate LQT1, LQT2 and LQT3 patients.
It was shown that it was possible to discriminate such patients using ECG data from only one lead, with the V6 and DI lead being the best with AUROC ranging from 0.63 to 0.70 for these leads. This indicates that the system trained with prolonged QT drug can discriminate between the different types of congenital LQT.
In summary, the above data shows that it is possible to train a machine classifier with ECG data, either composite or from single leads, and that the trained machine classifier can
| Number | Date | Country | Kind |
|---|---|---|---|
| 19305730.4 | Jun 2019 | EP | regional |
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/EP2020/065562 | 6/4/2020 | WO | 00 |
