Artificial intelligence for a cardiac drug study

Induced pluripotent cardiomyocyte A B S T R A C T Machine learning that is the vital part of modern artificial intelligence was used to classify the potential effects of two drugs on induced pluripotent stem cell-derived cardiomyocytes (iPSC-CM). Peak data were detected from calcium transient signals of cycling iPSC-CMs first at baseline, second exposed to adrenaline and third exposed with one of the drugs either flecainide or carvedilol. Various machine learning classification tests were executed to evaluate whether a drug did or did not affect signal shape after exposing adrenaline or the drugs. According to our classification results, machine learning can be applied to analyze possible effects of drugs on induced pluripotent stem cell-derived cardiomyocytes derived from the catecholaminergic polymorphic ventricular tachycardia patients ’ cell samples.


Introduction
Calcium cycling between sarcoplasm and sarcoplasmic reticulum (SR) is essential for the excitation-contraction coupling of cardiomyocytes.This movement from intracellular storage to sarcoplasm and back to the storage in SR can be visualized with calcium transients.Calcium is stored in the sarcoplasmic reticulum and influx of calcium through L-type Ca2+ channels triggers stored calcium to be released and bound to sarcomeres and induce cardiac contraction.Abnormal calcium cycling could promote severe cardiac arrhythmias in various cardiac disorders, e.g., in catecholaminergic polymorphic ventricular tachycardia (CPVT).Human-and patient-specific iPSC-derived cardiomyocytes are a great tool to study both normal human cardiomyocytes and diseased cells carrying various mutations.With these cells calcium cycling and transients can be analyzed in detail.
Machine learning for calcium transient signal analysis of induced pluripotent stem cell-derived cardiomyocytes was begun to separate different cardiological diseases from each other [1] before considering drugs.Thus far, research of drug influence on calcium transient signals of human induced pluripotent stem cell-derived cardiomyocytes is rather seldom carried out applying machine learning.The electrophysiological influence of chronotropic drugs was researched by using machine learning [2].An analysis of mechanistic action of cardiological drugs was performed with machine learning methods [3].Machine learning was applied to human induced pluripotent stem cell-based drug cardiotoxity testing [4].Induced pluripotent stem cell-based drug screening was reviewed for other areas in addition to cardiomyocytes [5].Furthermore, more widely machine learning was reviewed for several various drug-related topics [6] as well as for cell therapy [7].Clearly, machine learning has become meaningful in various ways in biomedicine and recently also for the research of calcium transient signals of human induced pluripotent stem cell-derived cardiomyocytes.
Recently, we have begun to research the use of machine learning for an analysis of drug effects where we first used calcium transient signals of induced pluripotent stem cell-derived cardiomyocytes that were exposed with adrenaline to generate arrhythmias [8].These induced pluripotent stem cell-derived cardiomyocytes carried mutations for genetic cardiac disease, catecholaminergic polymorphic ventricular tachycardia (CPVT).Next such cells were exposed with the drug called dantrolene and their response signals were divided into three classes by a biotechnology expert.The classes were responders, semi-responders and non-responders.
In the current machine learning research, we studied effects of two other antiarrhythmic drugs, flecainide and carvedilol, on abnormal calcium transient signals found with CPVT.New test arrangements were applied by categorizing the signals into only two classes of calcium transient signals, abnormal and normal.In addition, calcium transient signal data originated from induced pluripotent stem cell-derived cardiomyocytes of a healthy control subject (wild type, WT) were classified by comparison.Corresponded to Ref. [8], we now had a more interesting or precise test arrangement, because for every pair of binary classification there was the same cell used, i.e., first a baseline cell without any exposing and second the same cell exposed to adrenaline or first a cell exposed to adrenaline and second the same cell exposed to one of the two drugs.

Applied induced pluripotent stem cell-derived cardiomyocytes
The research was approved by the Ethics Committee of Pirkanmaa Hospital District as to culturing and differentiating of human induced pluripotent stem cell lines (R08070).The techniques related to human iPSC-CMs were given earlier [9].The iPSC cell lines used contained CPVT lines generated from four CPVT patients carrying mutations.Cell lines 03706CPVT, 05404CPVT, 05603CPVT and 05605CPVT were derived from patients with CPVT and cell line 04602WT from a healthy control.Patient-specific cell lines carry mutations in RYR2 gene: cell lines 05603CPVT and 05605CPVT are from the same individual and carry exon 3 deletion c.168-301_c.273 + 722del1128, cell line 03706CPVT carries point mutation p.L4115F (c.12343C > T), and cell line 05404CPVT carries point mutation p.V4653F (c.13957G > T) [9].
Antiarrhythmic evaluation of studied drugs was assessed by recording calcium transients at the baseline (perfusate only), 3min after 1 μM adrenaline incubation and if recorded transients exhibited calcium abnormalities, 5min after 0,25 μM carvedilol (Sigma) or 10 μM flecainide (Sigma) perfusion at the presence of 1 μM adrenaline.All the recordings were 30sec long with a sampling rate of 20 ms.
The calcium transients were categorized by an expert working with the cells and having several years expertise in analyzing the signals.The abnormal beating presented at least one of the following abnormalities: oscillation (OS), multiple peaks (MP), alternans (AL), plateau abnormality (PA), low peaks (LP) or irregular phase (IP) [9].Calcium transient was categorized as oscillation if there were more than three peaks in the same event, as multiple peaks if there were two or three peaks in the event, low peaks if the peak amplitude was at least 10% lower and 10% of the amplitude of a normal peak in the trace, plateau abnormality if there were significant disturbances in the rise or decay phase of a peak, alternans if there were regular alternating patterns in the amplitude of six or more consecutive peaks in the trace.If none of the abnormalities were observed, the beating was categorized normal.CPVT phenotype of the recorded cells was confirmed by the detection of adrenaline induced calcium abnormalities and only cells with arrhythmias after adrenergic stimulations were included in the drug studies.

Data
Earlier we have implemented a method in order to detect peaks from calcium transient signals and to compute their peak attributes [8,10].Peaks were detected from calcium transient signals according to computing first derivative applying short successive segments of a few sample values from the beginning of a calcium transient signal to its end.The signals had been sampled at 50 Hz.When first derivative values increased approximately from zero fast to positive, the beginning of a peak was detected.Next, decreasing first derivative values again close to zero, the peak maximum was encountered.After a rapid change to negative first derivative values again close to zero, the peak end was reached.However, small peaks less than 8% compared with the large amplitude peaks of a signal were dropped out as probable noise.
The biotechnology expert assessed whether the type of every calcium transient signal was either normal or abnormal (see 2. Material for classification).Normal signals are such that consist of sequential quite smooth peaks being approximately of the same amplitude, their left and right amplitudes are roughly of the same size, and peaks appear mostly in regular rhythm one after another.Instead, abnormal calcium transient signals are clearly irregular when not following the previous definitions, but including asymmetric peaks, peaks of clearly different sizes, and perhaps with considerably different intervals between peaks.
Fig. 1(a) shows an adrenaline exposure signal containing varying peak sizes and forms.Its response signal with regular peaks after drug exposure is in Fig. 1(b).Fig. 2(a) and (b) show a case where the drug decreased irregularity in the response signal, but only partially.
Altogether, there were 600 signals of four CPVT cell lines and signals of one WT cell line.Respectively, there were 20,071 and peaks.The numbers of peaks varied quite much between signals.Peak shapes and sizes varied between signals and inside abnormal signals.
The main data originated from four CPVT cell lines.One control type (WT) cell line was also included.There were calcium transient signals in four categories: baseline, adrenaline exposures and drug treatment after adrenaline.The baseline calcium transient signal of a cell originated from the same cell as its corresponding adrenaline transient signal.
Similarly, the cell of an adrenaline calcium transient signal was the same to its drug transient signal.Thus, an adrenaline transient signal had the corresponding drug transient signal from one of the drugs.In addition, there were also cells with adrenaline transient signals that had their corresponding baseline signals, but no corresponding drug transient signal.
In healthy CMs, adrenaline only increases the beating rate.In this current study, the WT iPSC-CMs signals at baseline were normal, but for the CPVT cell lines they were mostly abnormal.After drug treatment, calcium transient signals were either normal or abnormal depending on whether one of the drugs affected or did not.

Peak attributes of calcium transient signals
For machine learning computation, 14 peak attributes were computed for every detected valid peak of a signal, either baseline, adrenaline exposure or both drug exposures.The full details of the computation of the attributes are presented earlier [10][11][12].The peak attributes are shown in Fig. 3 as follows.These attributes formed the data of a peak of a calcium transient signal for each cell line used in the classifications between the different categories of calcium transient signals named above.

Classification methods
Classification was performed with such algorithms that we have seen to be the most powerful for various classification tasks of calcium transient signal data in our recent research [10,12].They were support vector machines with various kernels, random forests and nearest neighbor searching with different distance measures.
We tested four classification algorithms which were k-Nearest Neighbor classifier (kNN) [13], Classification and Regression Trees (CART) [14,15] called decision tree here, Random Forests (RF) [16], and Support Vector Machine (SVM) [8,17].With kNN, we varied the choice of distance measures, which were Chebychev, cityblock, correlation, cosine, Euclidean, Mahalanobis, standardized Euclidean, and Spearman.The weighting of instances was kept equal, and we tested the odd k values from 1 to 51.In the case of RF, we examined the number of trees in a forest ranging from 1 to 100 with a step size of 1.For SVM, we tested four different kernel functions, which were the linear, quadratic, polynomial kernel of degree 3, and the Radial Basis Function (RBF).Hyperplane optimization was made using ISDA algorithm [18].A regularization parameter, boxconstraint (C), is a common parameter for all kernels and for boxconstraint parameter the parameter value space tested was {2 − 9 , 2 − 8 , …, 2 10 }.Besides boxconstraint value, the RBF kernel requires the tuning of the kernel scale parameter and for this parameter the same parameter value space was used as for the boxconstraint.
The actual classification in all scenarios was performed with a similar procedure as in Ref. [8], i.e., leave-one-signal data-out approach (LOSDO).Before LOSDO, we made a z-score standardization for the whole dataset.In LOSDO, in each round we extract the peak-based data from the ith signal and leave it to a test set and the rest of the data forms the training set of the round in question.Then we train the classification algorithm using the specific parameter values if necessary and predict class labels for the test set instances.We take the mode of the test set peak predictions as a final prediction for the ith signal and use this information when evaluating the performance measures.If a classification algorithm requires the fine-tuning of hyperparameters, LOSDO procedure is repeated with all parameter value combinations and the best parameter value is selected based on the highest signal-level accuracy.
Besides signal-level accuracy (defined as a sum of diagonal elements of a confusion matrix divided by the sum of all elements in a confusion matrix), we evaluated the confusion matrix and true positive rates (TPs) (proportion of correctly classified signals in percentages) for each class as the performance measures in this study.Furthermore, we computed in which a and d are the numbers of correctly classified signals of two classes and b and c are the numbers of their incorrectly classified signals, respectively.Odds ratio is commonly used to estimate the strength of association between an exposure and an event.However, if some of those numbers is equal to zero, odds ratio cannot be calculated.

Results
We computed the machine learning classifications of four CPVT cell lines separately to see whether their results differ between the cell lines.For every cell line, we built seven computation types: baseline data vs. adrenaline exposures, normal vs. abnormal flecainide exposures, abnormal flecainide vs. adrenaline exposures, normal flecainide vs. adrenaline exposures, normal vs. abnormal carvedilol exposures, abnormal carvedilol vs. adrenaline exposures, and normal carvedilol vs. adrenaline exposures.These were tested with kNN applying different distance measure, CART decision tree, RF and SVMs with linear, quadratic, cubic and RBF kernels.
Table 1 presents the results of the first CPVT iPSC-CMs in flecainide test.For the normal vs. abnormal signals, the highest classification accuracy of 78 % (in Bold) was given by SVM with either linear or RBF kernel.However, they did not identify the smaller class of the normal  best result, which is high, even if both are abnormal, obviously, somehow in different ways.For the adrenaline vs. normal signals after the drug, nearest neighbor searching with city block metric, random forest and SVM with linear and quadratic kernels were the best with accuracy 88 %.For the second and third classification set-ups odd ratios 26.5 and 55.0 show the best classifiers.Odds ratio is more sensitive than accuracy because the former contains multiplication and division, whereas the latter addition and division.
Table 2 shows the results of the first CPVT iPSC-CMs in carvedilol test.Again, the number of the abnormal signals not affected by the drug was considerably greater than that of normal signals.The normal signals were classified rather poorly, but better than those in Table 1.The best classification accuracy of 83 % was gained by nearest neighbor searching with cosine measure and SVM with RBF.Nearest neighbor searching with Spearman measure was close to the best accuracy of 74 % generated by SVM with linear kernel for adrenaline signals vs. abnormal after carvedilol drug.For adrenaline vs. normal signals after the drug, nearest neighbor searching with correlation measure reached the highest classification accuracy of 80 %.Odds ratios 20.2, 8.4, and 16.0 also indicated the best classifiers of three classification set-ups.
We also tested the set-up of the baseline signals vs. adrenaline signals.The results are not presented in the table when the set-up can be seen less important than those three shown in the tables.The best classification accuracy of 57 % was given by SVM with linear kernel and SVM with RBF kernel.The latter gave the true positive rates of 40 % for the adrenaline signals and 74 % for the baseline signals.There were signals in both signal sets.As mentioned above, all adrenaline signals were abnormal.For this cell line, all baseline signals occurred to be also such.Obviously, this explains somewhat low classification accuracy.
In Table 3 there are the results of the second CPVT patient-specific iPSC-CMs with flecainide.For the normal vs. abnormal signals after the flecainide exposure, random forest gained the highest accuracy of %.For the adrenaline signals vs. abnormal after flecainide drug, nearest neighbor searching with correlation measure gave the highest accuracy side absolute minimum of the first derivative, locations for (g) right side absolute minimum and (h) maximum of the second derivative, (i) peak surface area between the peak curve and the line from the peak beginning to the end, (j) interval from the peak maximum back to the maximum of the preceding peak or to the beginning of the signal for its first peak, (k) interval from the peak beginning to the place of the first derivative maximum, (l) interval from the peak maximum to the place of the first derivative absolute minimum, (m) interval for the mean of the halves of the left and right sides amplitudes, and (n) peak curve length. of 78 %.Again, nearest neighbor searching with Chebychev metric produced the best accuracy of 84 % in the test set-up of the adrenaline vs. normal signals after the drug.Odds ratios 7.9, 14.0, and 50.4 also gave the same best classifiers.They emphasize the differences between three classification set-ups, whereas their accuracies differ less form each other.
The peak data of Table 3 were computed with t-Distributed Stochastic Neighbor Embedding (t-SNE) and shown in Fig. 4 for three classification set-ups.In all these three, both classes are separated relatively well as the accuracies and odds ratios of the three classification set-ups denote in Table 3.There are more peaks in the class of the abnormal signals after adrenaline exposure than in those of flecainide exposures.This is quite typical because adrenaline generates frequent peaks as seen in Figs. 1 and 2. On the contrary, the drug can decrease the number of peaks, particularly, if the drug affects.Table 5 contains the flecainide results of the third CPVT specific iPSC-CMs.For the normal vs. abnormal signals with flecainide, SVM with linear kernel produced the highest classification accuracy of 78 %.Nevertheless, all the numbers of the signals were exceptionally small meaning that the results of the table are only indicative.For the adrenaline vs. abnormal with flecainide, SVM with linear kernel was able to achieve the best accuracy of 63 %.For the adrenaline vs. normal signals with flecainide, nearest neighbor searching with four different distance measures, SVM with linear kernel as well as with cubic kernel gave the highest accuracy of 90 %.In the second classification set-up odds ratio of 3.0 only was possible to calculate to support the accuracy result.
In Table 6  As to the baseline signals vs. adrenal signals, there were 46 signals in both sets.A great majority of the baseline signals were abnormal.The best accuracy of 54 % was given by SVM with linear kernel or RBF kernel.Both true positive rates of SVM with linear kernel were also %.
To observe how the used peak data are distributed in the feature space we computed two-dimensional spaces with t-Distributed Stochastic Neighbor Embedding (t-SNE).Fig. 5 depicts the peak data associated with three classification set-ups of Table 8.In Fig. 5 two classes differ well from each other.Nonetheless, the highest accuracy and odds ratio are better in the first set-up than in the second and third set-ups.
Ultimately, the results of the fifth iPSC CMs that were control cells (WT) are presented.Notice that all its signals with adrenaline exposure were normal differing from those of CPVT cell lines.For the normal vs. abnormal signals of flecainide, nearest neighbor searching with Spearman measure and SVM with RBF kernel gave the highest accuracy of 83 % in Table 9.For the normal adrenaline vs. abnormal signals with In Table 10, associated with the normal vs. abnormal signals of carvedilol, random forests achieved the greatest accuracy of 86 %.Normal signals were quite difficult to separate perhaps for the sake of their small number.For the normal adrenaline vs. abnormal signals with carvedilol, SVM with linear kernel produced the highest accuracy of 86 %.For the normal adrenaline vs. normal carvedilol signals, SVM with the linear kernel attained the greatest accuracy of 67 %.The best odds ratios 12.8 and 4.0 are those of the first and third classification set-ups.
For to the baseline signals vs. adrenal signals, there were 100 signals in both sets.The greatest accuracy of 56 % was given by SVM with the quadratic kernel.The true positive rate of adrenaline was then 68 % and that of baseline 44 %.

Discussion and conclusion
iPSC-CMs have been demonstrated by us [19] and by others [20] to reproduce the drug response observed with the patient donating the primary cells.We have previously successfully demonstrated that machine learning can identify disease and even mutation specific iPSC-CMs from each other [8].In the current study we wanted to analyze this in more detail with drugs currently used for the treatment of this genetic arrhythmia, CPVT.Here we also wanted to compare the full response to limited or no response.
For cell line 03706CPVT (mutation), the best classification accuracies were observed when abnormal adrenaline vs. abnormal with the drug and abnormal adrenaline vs. normal with the drug were compared (Tables 1 and 2).The results, however, show that there were differences between these two classes in both binary classifications.The best classification accuracies of normal vs. abnormal with the drug, were also high, but less than in the other comparisons.Perhaps due to the small number of cells in these groups (12 and 10 signals in normal compared to abnormal, respectively) the calculations gave low true positive rates.
For 05404CPVT iPSC-CMs (mutation), the best classification   3 and 4 Those of the first tests (normal to abnormal) were good and the true positive values of the normal better than with the first cell line, however, mainly below 50% perhaps again for sake of the disparity of the normal and abnormal signals.
For 05603CPVT iPSC-CMs (mutation), the greatest accuracy values are high in Tables 5 and 6 Nevertheless, the very low numbers of the signals in Table 5 show only guiding results.All best classification accuracies and true positive values are high.
For 05605CPVT iPSC-CMs (mutation), all greatest classification accuracies and true positive values are high in Tables 7 and 8 Although the numbers of the signals in the subset are rather low, their ratios are of approximately equal size.
For control cells, 04602WT iPSC.CMs, all best classification accuracies and true positive values except for the abnormal of the first tests in Table 10 are very high.Once again, the disparity of the numbers of the signals was present in the first tests.
In conclusion, these classification results indicate that there were clear differences between the normal and abnormal signals after the  8. cells first exposed with adrenaline and then with the drug.There were also differences between cells first exposed with adrenaline, then with the drug, when in the classifications these latter were divided into either normal or abnormal signals based on types with or without the drug influence.The test set-ups were paired in a sense that there was always the same cell first exposed with adrenaline and second with the drug, which was performed for several cells.Similarly, the test was first made for baseline cell and for the same cell after its adrenaline exposure.The calcium transition signal data of four CPVT cell lines and one WT cell line were classified with several machine learning methods.The best classification accuracies and true positive rates attained were high, once equal to 60 % and typically much higher from 74 % to 90 % and up to the maximum of 96 % so that a clear majority of each test arrangement was then classified correctly.The odds ratios computed typically indicated the same best classifications.When odds ratio is more sensitive than accuracy, it emphasizes even a small difference between the results of various classification methods.
A data set should contain enough signals for modelling, more than 50 for both adrenaline exposure signals and their responding drug exposure signals.It would be interesting to continue our research with larger data sets than here.In addition, the "partial similarity or dissimilarity" of normal and abnormal exposure signals as seen in Figs.1(a It was not necessary and possible to consider manually individual peaks for the sake of their great number.Still, we saw this manual initial consideration of the entire signals inevitable to determine which are normal or abnormal signals, since there do not yet exist any general and precise definitions for abnormal and normal calcium transient peaks and signals that would be needed for the fully automatic process.It is important to notice that the patient specific iPSCs have been derived from patients having various mutations and thus when we study how well the calcium transient signals of diseased cell lines can be classified, this means that the individual patients could be classified to the correct disease category.In this paper we also demonstrate the drug effects in diseased cells can be analyzed with machine learning and we have previously demonstrated that patients and their iPSC-derived cardiomyocytes respond the same way on drug treatment [19].The classification for the separation of normal and abnormal signals was not the only target.The final successful target was to show that machine learning methods can also be applied to detect whether the drug affected.Altogether, the results attained show that machine learning is a very promising methodology for the research of the drug effects of induced pluripotent stem cell-derived cardiomyocytes.

Ethical statement
The research was approved by the Ethics Committee of Pirkanmaa Hospital District in association with Tampere University and Tampere University Hospital as to culturing and differentiating of human induced pluripotent stem cell lines (R08070).

Declaration of competing interest
None declared.

Fig. 1 .
Fig. 1.(a) A calcium transient signal of iPSC cardiomyocyte after a cell was exposed with adrenaline which generated as if arrhythmic peaks.Detected regular peaks are marked with green at their maxima and irregular peaks with other colors.(b) Thereafter, the same cell was exposed with flecainide which affected the cell and produced a relatively regular, i.e., normal signal.

Fig. 2 .
Fig. 2. (a)A calcium transient signal of another iPSC cardiomyocyte after a cell was exposed with adrenaline.(b) The same cell was then exposed with flecainide which did not affect positively enough, but gave an irregular, i.e., abnormal signal.

Fig. 3 .
Fig. 3. Peak attributes: (a) The left and (b) right side amplitudes of a peak, (c) left and (d) right side durations, locations for (e) left side maximum and (f) rightside absolute minimum of the first derivative, locations for (g) right side absolute minimum and (h) maximum of the second derivative, (i) peak surface area between the peak curve and the line from the peak beginning to the end, (j) interval from the peak maximum back to the maximum of the preceding peak or to the beginning of the signal for its first peak, (k) interval from the peak beginning to the place of the first derivative maximum, (l) interval from the peak maximum to the place of the first derivative absolute minimum, (m) interval for the mean of the halves of the left and right sides amplitudes, and (n) peak curve length.

Fig. 4 .
Fig. 4. (a) The t-SNE distributions for two classes of the normal vs. abnormal with flecainide, (b) abnormal adrenaline vs. abnormal with flecainide, and (c) abnormal adrenaline vs. normal with flecainide peak data used to compute the results of Table3.

Fig. 5 .
Fig. 5. (a) The t-SNE distributions for two classes of the normal vs. abnormal with carvedilol, (b) abnormal adrenaline vs. abnormal with carvedilol, and (c) abnormal adrenaline vs. normal with carvedilol peak data used to compute the results in Table8.
) and Fig.2(b) could be studied.It should be noticed that at the beginning of the research the biotechnology expert considered the data by dividing the entire signals into either normal or abnormal cases.

Table 1
Cell line 03706CPVT of calcium transient signals when cells exposed with adrenaline and then drug given: (1) Flecainide changed 12 to normal vs. 43 that stayed abnormal, (2) 43 abnormal adrenaline signals vs. 43 that stayed abnormal, and (3) 12 abnormal adrenaline signals vs. 12 that were converted to normal signals (total 110 signals).Ac classification accuracy, TP n true positive rate of normal calcium transient signals, TP b that of abnormal signals, TP d that of abnormal adrenaline exposure signals, and odds ratio OR.

Table 2
Cell line 03706CPVT of calcium transient signals when cells exposed with adrenaline and then drug given: (1) carvedilol changed 10 to normal vs. 31 that stayed abnormal, (2) 31 abnormal adrenaline signals vs. 31 that stayed abnormal, and (3) 10 abnormal adrenaline signals vs. 10 that were converted to normal signals (a total of 82 signals).Ac classification accuracy, TP n true positive rate of normal calcium transient signals, TP b that of abnormal signals, TP d that of abnormal adrenaline exposure signals, and OR odds ratio.

Table 4
consists of the results of the second CPVT patient-specific iPSC-CMs for the carvedilol tests.For the normal vs. abnormal signals after the carvedilol exposure, nearest neighbor searching with cityblock metric gained the highest accuracy of 79 %.For the adrenaline signals vs. abnormal signals after carvedilol drug, SVM with linear kernel gave the highest accuracy of 60 %.For the adrenaline vs. normal signals, the best was SVM with RBF kernel reaching 78 %.Odds ratios 8.8, 2.4, and

Table 3
Cell line 05404CPVT of calcium transient signals when cells exposed with adrenaline and then drug given: (1) flecainide changed 19 to normal vs. 39 that stayed abnormal, (2) 39 abnormal adrenaline signals vs. 39 that stayed abnormal, and (3) 19 abnormal adrenaline signals vs. 19 that were converted to normal signals (a total of 116 signals).Ac classification accuracy, TP n true positive rate of normal calcium transient signals, TP b that of abnormal signals, TP d that of abnormal adrenaline exposure signals, and OR odds ratio.
there are the results of the third CPVT specific iPSC-CMs subject with carvedilol drug.Nearest neighbor searching with three different distance measures, random forest, and SVM with quadratic and RBF kernels gave the best accuracy of 83 %.Subject to the adrenaline signals vs. abnormal with carvedilol, SVM with linear kernel generated the best accuracy of 85 %.For the adrenaline vs. normal signals, random forest and SVM with linear and RBF kernels reached the best accuracy of 72 %.Again, odds ratios 27.3, 31.7, and 7.9 expressed the most effective classification methods.Regarding the baseline signals vs. abnormal adrenaline signals, there were 43 signals in both sets.The most effective classifier was SVM with RBF kernel with the accuracy of 59 %.It generated the true positive rates of 70 % and 49 % for the adrenaline and baseline signals.Almost all baseline signals were abnormal which may have complicated classification process if the signals of two sets resembled each other.In Table 7, the results of the fourth CPVT patient-specific iPSC-CMs with flecainide drug are presented.For the normal vs. abnormal signals after flecainide exposure, SVM with RBF kernel gained the highest accuracy of 74 %.As regards the adrenaline signals vs. abnormal signals with flecainide, nearest neighbor searching with correlation measure and SVM with linear kernel attained the highest classification accuracy of 84 %.As for the adrenaline vs. normal signals, nearest neighbor searching with either correlation or Spearman measure, random forest and SVM with RBF kernel achieved the best accuracy of 87 %.Odds ratios 8.7 and 42.3 of the first and third classification are the highest values.In Table 8, there are the results of the fourth CPVT patient-specific iPSC-CMs for the carvedilol drug.Regarding the normal vs. abnormal signals with carvedilol, nearest neighbor searching with Mahalanobis metric, random forest and SVM with RBF kernel attained the best accuracy of 79 %.As to the adrenaline vs. abnormal signals, nearest neighbor searching with Chebychev metric gained the best accuracy of 69 %.As for the adrenaline vs. normal signals, SVM with either quadratic or RBF kernel attained the highest accuracy of 72 %.The highest odds ratios are 16.5, 5.1, and 9.0.

Table 4
Cell line 05404CPVT of calcium transient signals when cells exposed with adrenaline and then drug given: (1) carvedilol changed 9 to normal vs. 24 that stayed abnormal, (2) 24 abnormal adrenaline signals vs. 24 that stayed abnormal, and (3) 9 abnormal adrenaline signals vs. 9 that were converted to normal signals (a total of 66 signals).Ac classification accuracy, TP n true positive rate of normal calcium transient signals, TP b that of abnormal signals, TP d that of abnormal adrenaline exposure signals, and OR odds ratio.

Table 5
Cell line 05603CPVT of calcium transient signals when cells exposed with adrenaline and then drug given: (1) flecainide changed 5 to normal vs. 4 that stayed abnormal, (2) 4 abnormal adrenaline signals vs. 4 that stayed abnormal, and (3) 5 abnormal adrenaline signals vs. 5 that were converted to normal signals (a total of signals).Ac classification accuracy, TP n true positive rate of normal calcium transient signals, TP b that of abnormal signals, TP d that of abnormal adrenaline exposure signals, and OR odds ratio., SVM with linear and RBF kernels attained the best accuracy of 94 %.For the normal adrenaline vs. normal flecainide signals, SVM with RBF kernel reached the greatest accuracy of all, 96 %.The very high odds ratios of 22.9, 385, and 676 indicate good classification results. flecainide

Table 6
Cell line 05603CPVT of calcium transient signals when cells exposed with adrenaline and then drug given: (1) carvedilol changed 18 to normal vs. 23 that stayed abnormal, (2) 23 abnormal adrenaline signals vs. 23 that stayed abnormal, and (3) 18 abnormal adrenaline signals vs. 18 that were converted to normal signals (a total of 82 signals).Ac classification accuracy, TP n true positive rate of normal calcium transient signals, TP b that of abnormal signals, TP d that of abnormal adrenaline exposure signals, and OR odds ratio.

Table 7
Cell line 05605CPVT of calcium transient signals when cells exposed with adrenaline and then drug given: (1) flecainide changed 15 to normal vs. 16 that stayed abnormal, (2) 16 abnormal adrenaline signals vs. 16 that stayed abnormal, and (3) 15 abnormal adrenaline signals vs. 15 that were converted to normal signals (a total of 62 signals).Ac classification accuracy, TP n true positive rate of normal calcium transient signals, TP b that of abnormal signals, TP d that of abnormal adrenaline exposure signals, and OR odds ratio.

Table 8
Cell line 05605CPVT of calcium transient signals when cells exposed with adrenaline and then drug given: (1) carvedilol changed 17 to normal vs. 12 that stayed abnormal, (2) 12 abnormal adrenaline signals vs. 12 that stayed abnormal, and (3) 17 abnormal adrenaline signals vs. 17 that were converted to normal signals (a total of 58 signals).Ac classification accuracy, TP n true positive rate of normal calcium transient signals, TP b that of abnormal signals, TP d that of abnormal adrenaline exposure signals, and OR odds ratio.

Table 9
Cell line 04602WT of calcium transient signals when cells exposed with adrenaline and then drug given: (1) flecainide stayed 27 normal vs. 36 that changed to abnormal, (2) 36 normal adrenaline signals vs. 36 that changed to abnormal, and (3) 27 normal adrenaline signals vs. 27 that stayed normal signals (a total of 126 signals).Ac classification accuracy, TP n true positive rate of normal calcium transient signals, TP b that of abnormal signals, TP d that of abnormal adrenaline exposure signals, and OR odds ratio.

Table 10
Cell line 04602WT of calcium transient signals when cells exposed with adrenaline and then drug given: (1) carvedilol stayed 37 normal vs. 7 that changed to abnormal, (2) 7 normal adrenaline signals vs. 7 that changed to abnormal, and (3) 37 normal adrenaline signals vs. 37 that stayed normal signals (a total of 88 signals).Ac classification accuracy, TP n true positive rate of normal calcium transient signals, TP b that of abnormal signals, TP d that of abnormal adrenaline exposure signals, and OR odds ratio.Normal vs. abnormal with carvedilol Abnormal adrenaline vs. abnormal with carvedilol Abnormal adrenaline vs. normal with carvedilol