TELK OMNIKA T elecommunication, Computing, Electr onics and Contr ol V ol. 24, No. 1, February 2026, pp. 142 ∼ 150 ISSN: 1693-6930, DOI: 10.12928/TELK OMNIKA.v24i1.27445 ❒ 142 R WT -Net: A h ybrid ResNet-wa v elet-transf ormer f or early detection of left v entricular h ypertr oph y Hoang Huu T o Nguy en 1 , Phuong Huu Nghia Le 2 , Lam Mai 1 , Nguy en Pham Ho T r ong 3 1 F aculty of Computer Science, V ietnam-K orea Uni v ersity of Information and Communication T echnology , The Uni v ersity of Da Nang, Da Nang, V iet Nam 2 Department of Biomedical Clinical Equipment T echnology , San Jacinto Colle ge, Houston, T e xas, United States of America 3 Department of Information Assurance, FPT Uni v ersity , Da Nang, V iet Nam Article Inf o Article history: Recei v ed Aug 2, 2025 Re vised Oct 29, 2025 Accepted Dec 8, 2025 K eyw ords: Deep learning Electrocardiogram Left v entricular h ypertroph y R WT -Net T ransformer W a v elet transform ABSTRA CT Early detection of left v entricular h ypertroph y (L VH), a k e y predictor of heart f ailure and strok e, is critical. Ho we v er , standard 12-lead electrocardiogram (ECG) criteria suf fer from lo w sensiti vity . While deep learning sho ws promise, a research g ap e xists for models that rob ustly inte grate di v erse si gnal fea- tures to impro v e detection, especially sensiti vity . W e propose ResNet-w a v elet- transformer net (R WT -Net), a h ybrid architecture that fuses deep morpholog- ical features from a ResNet1D with statistical t ime-frequenc y features from a w a v elet pack et transform (WPT) using a transformer encoder . The model w as e v aluated on the PTB-XL dataset (11,201 recordings) using a stringent, patient- le v el 5-fold cross-v alidation. R WT -Net achie v ed a mean area under the curv e (A UC) of 0.9868 and F1-score of 0.8725. Critically , its w a v elet-enhanced stream yielded signicantly higher sensiti vity compare d to a ResNet-transformer base- line (0.8964 vs. 0.8716, p=0.0039), better addressing the clinical need to mini- mize f alse ne g ati v es. A k e y limitation is the reliance on ECG-based labels, not an ec hocardiograph y gold standard. R WT -Net demonstrates potential as a re- liable, automated screening tool to prioritize at-risk patients for further clinical assessment. This is an open access article under the CC BY -SA license . Corresponding A uthor: Nguyen Pham Ho T rong Department of Information Assurance, FPT Uni v ersity FPT City , Ngu Hanh Son W ard, Da Nang City , V iet Nam Email: nguyenpht@fe.edu.vn 1. INTR ODUCTION Left v entricular h ypertroph y (L VH), the thick ening of the heart’ s main pumping chamber , is a signi- cant cardio v ascular abnormality often de v eloping in response to pressure o v erload conditions lik e h ypertension. As a strong independent predictor of adv erse e v ents such as heart f ailure and strok e, its early and accurate de- tection is paramount for timely interv ention [1]. F or decades, the 12-lead electrocardiogram (ECG) has been the primary non-in v asi v e screening tool, yet traditional v oltage-based criteria lik e Sok olo w-L yon suf fer from critically lo w sensiti vity , often detecting less than 25% of true L VH cases. This diagnostic g ap underscores an ur gent need for more rob ust screening methods. Articial intelligence (AI), particularly deep learning, of fers a po werful alternati v e by learning com- ple x patterns from ECGs [2]-[4]. Con v olutional neural netw orks (CNNs), especially ResNet architectures, ha v e specically demonstrated success in e xtracting morphological features for L VH detection [5]-[8], with J ournal homepage: https://telk omnika.uad.ac.id/inde x.php/TELK OMNIKA Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control ❒ 143 meta-analyses conrming their high diagnostic accurac y [9], [10]. Concurrently , interest has gro wn in h ybrid architectures combining CNNs with sequence models [11] or fusing deep features with handcrafted ones, such as those deri v ed from w a v elet transforms [12]-[14]. Despite these adv ancements, a signicant methodological g ap persists in the literature. Man y studi es emplo y simplistic data splitting strate gies (e.g., at the record le v el), which risk data leakage from patients with multiple recordings, leading to inated and unre liable performance metrics. A clinically viable model must be v alidated with stringent, patient-le v el data separation to ensure it can generalize to unseen indi viduals. This study addresses this g ap by introducing ResNet-w a v elet-t ransformer net (R WT -Net), a no v el h ybrid architec- ture e v alua ted under a rigorous, statistically rob ust frame w ork. Our contrib utions are: (i) the proposal of the R WT -Net architecture with a sophisticated fusion mechanism that inte grates deep morphological features with handcrafted statistical w a v elet features; (ii) the implementation of a stringent patient-stratied 5-fold cross- v alidation protocol, repeated across multiple random seeds, to ensure reliable and generalizable results; and (iii) a comprehensi v e ablation and statistical analysis that v alidates the contrib ution of each model component, re v ealing the critical role of w a v elet features in enhancing sensiti vity for L VH detection, a k e y requirement for ef fecti v e clinical screening. 2. METHOD 2.1. Dataset and pr epr ocessing The study utilized the PTB-XL dataset, a lar ge, publicly a v ailable electrocardiograph y dataset con- taining 21,837 clinical 12-lead ECGs from 18,885 patients [15], [16] an ecosystem that also includes compre- hensi v e feature sets for traditional machine learning [17]. F or this binary classication task, a specic cohort w as curated based on the pro vided PTB-XL diagnostic statements. Records were included as L VH-positi v e if the L VH diagnostic statement w as present, re g ardless of other co-e xisting conditions (e.g., MI and ISCA). Records were selected as the ne g ati v e class only if their sole diagnostic superclass w as normal (NORM). This resulted in a nal study cohort of 11,201 recordings from 8,960 unique patients. It is a critical limitation of this study that all labels are based on the pro vided ECG annotations from the PTB-XL dataset, not on a clinical gold standard such as echocardiograph y , which may introduce diagnostic uncertainty . Prior to model training, each 12-lead signal, sampled at 100 Hz for 10 seconds (1000 times teps), underwent a tw o-stage preprocessing pipeline. First, a second-order Butterw orth band-pass lter with cutof f frequencies of 0.5 Hz and 45 Hz w as applied to m itig ate common ECG artif acts. Second, each lead w as independently standardized using Z-score normalization as sho wn in (1), where x is the ra w signal, µ is the mean, and σ is the standard de viation. This ensures all leads are on a comparable scale, which is crucial for impro ving model con v er gence. x nor mal iz ed = x − µ σ (1) The qualitati v e impro v ement from this pipeline is visualized in Figure 1(a). P anel (a) displays the ra w signals, while panel (b) sho ws the corresponding signals after ltering and normalization, demonstrating the ef fecti v e remo v al of baseline w ander . 2.2. Pr oposed method: R WT -Net ar chitectur e Our proposed model, R WT -Net, is a h ybrid, dual-stream architecture designed to create a holistic s ig- nal representation by inte grating features from distinct yet complementary domains. The rst stream focuses on e xtracting high-le v el morphological patterns using a ResNet-based deep encoder , while the second stream captures statistical time-frequenc y characteristics via w a v elet pack et transform. By fusing these tw o informa- tion sources through a transformer encoder , the model ef fecti v ely le v erages both automated feature learning and domain-specic signal processing kno wledge. 2.2.1. W a v elet featur e extraction str eam In parallel with the time-domain stream, the ra w 12-lead ECG signal is processed by a w a v elet pack et transform (WPT) block using a Daubechies 4 (db4) w a v elet with a decomposition le v el of 3. This decomposes the signal into 2 3 = 8 frequenc y sub-bands (terminal nodes). From each node’ s coef cients ( c j ), four statistical features are calculated: ener gy , standard de viation, sk e wness, and kurtosis. These features from all 12 leads are R WT -Net: A hybrid ResNet-wavelet-tr ansformer for early detection of left ventricular (Hoang Huu T o Nguyen) Evaluation Warning : The document was created with Spire.PDF for Python.
144 ❒ ISSN: 1693-6930 then a v eraged, resulting in a rich, 32-dimensional handcrafted fea ture v ector ( f w av el et ∈ R 32 ) that encapsulates the signal’ s time-frequenc y characteristics, as sho wn in Figure 1(b). (a) (b) Figure 1. Preprocessing and feature e xtraction visualizations: (a) ra w and preprocessed ECG signals and (b) e xample 32-dimensional WPT feature v ector 2.2.2. ResNet-transf ormer backbone The preprocessed time-domain signal is fed into a deep feature encoder based on a ResNet1D archi- tecture. The core of this netw ork relies on residual blocks to help pre v ent v anishing gradients in deep netw orks. TELK OMNIKA T elecommun Comput El Control, V ol. 24, No. 1, February 2026: 142–150 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control ❒ 145 The learned feature maps are then projected to match the transformer’ s internal dimension ( d model ). The 32D w a v elet feature v ector is e xpanded and concatenated to the feature v ector at each time step. This fused repre- sentation is then fed into a transformer encoder , which uses a multi-head self-attention mechanism to capture comple x interdependencies. The output sequence is aggre g ated by a global a v erage pooling (GAP) layer and passed to an multilayer perceptron (MLP) head for nal classication. The o v erall architecture, which inte- grates these streams, is depicted in Figure 2. Figure 2. Ov erall R WT -Net frame w ork inte grating preprocessing, dual-stream (ResNet+W a v elet) architecture with transformer fusion, and e v aluation stages 2.3. Experimental setup 2.3.1. Ev aluation pr otocol T o ensure a rob ust and clinically rele v ant e v aluation, we emplo yed a str atied 5-fol d cross-v al idation scheme grouped by patient ID. This protocol guarantees that all recordings from a single patient belong e xclu- si v ely to either the training or the test set in an y gi v en fold, pre v enting data leakage. T o assess stability , the entire 5-fold process w as repeated four times with dif ferent random seeds (42, 52, 62, 72). All reported results are the mean and standard de viation o v er these repeated runs. 2.3.2. Baseline and ablation models T o conte xtualize performance, we compared R WT -Net ag ainst se v eral models: (a) XGBoost on 398 handcrafted tw o-stream superheterodyne free ele ctron lasers (TSFEL) features [18]; (b) 1D-CNN; (c) long short-term memory (LSTM); and (d) ablation models (ResNet-only , transformer -only , and R T -Net without w a v elets) to isolate each component’ s contrib ution. 2.3.3. Implementation details Models were trained on an NVIDIA T4 GPU using the AdamW optimizer [19] and a weighted cross- entrop y loss to handle class imbalance. A cosine anneali ng learning rate scheduler and early stopping (pa- tience=20) were used. Data augmentati on (Gaussian noise and time shifts) w as applied during training. K e y h y- perparameters are in T able 1. The implementation le v erages common scientic computing libraries in Python, such as Scikit-learn for e v aluation metrics [20]. R WT -Net: A hybrid ResNet-wavelet-tr ansformer for early detection of left ventricular (Hoang Huu T o Nguyen) Evaluation Warning : The document was created with Spire.PDF for Python.
146 ❒ ISSN: 1693-6930 T able 1. K e y h yperparameters for all e v aluated models Model P arameter V alue R WT -Net/R T -Net ResNet lters [32, 64, 128, 256] T ransformer layers/heads 2 / 4 d model / d f f 128 / 256 Optimizer (LR / weight decay) AdamW (2e-4 / 5e-3) XGBoost n estimators / max depth 1000 / 5 learning rate / early stopping 0.005 / 50 LSTM Hidden size / Num layers 128 / 2 Optimizer (LR / weight decay) AdamW (1e-4 / 1e-3) 1D-CNN Architecture 3 Con v layers, 2 FC layers Optimizer (LR / weight decay) AdamW (1e-4 / 5e-4) 2.4. Ev aluation metrics and statistical analysis Model performance w as e v aluated using area under the curv e (A UC), F1-score, accurac y , s ensiti vity , and specicity . Results are reported as mean ± standard de viation. An independent tw o-sample t-test (alpha = 0.05) w as used to determine if performance dif ferences were statistically signicant. 3. RESUL TS AND DISCUSSION 3.1. Ov erall perf ormance comparison The comparati v e results are summarized in T able 2 and illustrated in Figure 3. T o visualize the perfor - mance g ains across dif ferent architectures, Figure 3(a) present s a comparati v e bar chart of F1-scores and A UC v alues. Furthermore, Figure 3(b) pro vides a detailed vie w of the ablation results to emphasize the impact of each component. These visual comparisons are complemented by the recei v er operating characteristic (R OC) curv es in Figure 4(a), which conrm the superior and stable discriminati v e po wer of the ResNet-transformer based models. R WT -Net achie v ed the highest mean F1-score (0.8725 ± 0.0154) and A UC (0.9868 ± 0.0020). This performance is statistically superior to the ne xt best bas eline, XGBoost (F1-score 0.8518, p < 0 . 001 ), and f ar e xceeds the standard deep learning base lines 1D-CNN (F1-score 0.8058, p < 0 . 001 ) and LSTM (F1-score 0.7407, p < 0 . 001 ). T able 2. Ov erall performance comparison on the PTB-XL dataset (mean ± std o v er 4 runs) Model A UC F1-score (L VH) Sensiti vity Specicity Accurac y R WT -Net 0.9868 ± 0.0020 0.8725 ± 0.0154 0.8964 ± 0.0292 0.9629 ± 0.0109 0.9502 ± 0.0062 R T -Net 0.9865 ± 0.0022 0.8771 ± 0.0110 0.8716 ± 0.0188 0.9728 ± 0.0048 0.9535 ± 0.0044 XGBoost 0.9809 ± 0.0031 0.8518 ± 0.0112 0.8175 ± 0.0157 0.9762 ± 0.0034 0.9459 ± 0.0032 1D-CNN 0.9535 ± 0.0040 0.8058 ± 0.0111 0.7626 ± 0.0175 0.9694 ± 0.0046 0.9300 ± 0.0051 LSTM 0.9092 ± 0.0226 0.7407 ± 0.0381 0.6570 ± 0.0476 0.9728 ± 0.0073 0.9128 ± 0.0114 (a) (b) Figure 3. Model performance e v aluation, error bars represent standard de viation across runs: (a) o v erall model performance comparison and (b) ablation study performance comparison TELK OMNIKA T elecommun Comput El Control, V ol. 24, No. 1, February 2026: 142–150 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control ❒ 147 3.2. Ablation study: the clinical v alue of wa v elet featur es The ablation study , detailed in T able 3 and illustrated in Figure 3(b), systematically deconstructs the architecture. This visualization highlights ho w combining ResNet and transformer modules yields a dramatic performance increase compared to standalone congurations. The training dynamics and o v erall stability of these congurations are further documented in Figure 4. Specically , the R OC curv e comparison is sho wn in Figure 4(a), while the training and v alidation loss curv es are pro vided in Figure 4(b). The standalone ResNet-only and transformer -only models sho w respectable b ut limited perf ormance. Combining them in the R T -Net model yields a dramatic performance increase, demonstrating a po werful syn- er gy where ResNet e xtracts local features which the transformer conte xtualizes. While their o v erall A UC and F1-scores are statistically indistinguishable ( p > 0 . 05 ), a crucial trade-of f emer ges. The inclusion of the w a v elet stream in R WT -Net leads to a statistically signicant increase in sensiti vity (0.8964 vs. 0.8716, p=0.0039). This comes at the cost of a small b ut statistically signicant decrease i n specicity (0.9629 vs. 0.9728, p=0.0012). This trade-of f is highly desirable for a clinical screening tool, where prioritizing the detec- tion of true cases (high sensiti vity) is more critical than minimizing f alse alarms. T able 3. Ablation study results (mean ± std) Model conguration A UC F1-score (L VH) Sensiti vity ResNet only 0.9423 ± 0.0039 0.7834 ± 0.0129 0.7427 ± 0.0168 T ransformer only 0.9416 ± 0.0048 0.7747 ± 0.0125 0.7489 ± 0.0313 R T -Net (ResNet+transformer) 0.9865 ± 0.0022 0.8771 ± 0.0110 0.8716 ± 0.0188 R WT -Net (full model) 0.9868 ± 0.0020 0.8725 ± 0.0154 0.8964 ± 0.0292 3.3. P erf ormance analysis and inter pr etability The training process, as visualized through the loss curv es in Figure 4(b), sho ws st able con v er gence without signicant o v ertting. T o enhance clinical trust and pro vide insight into the model’ s decision-making process, a critical aspect of medical AI [21], we emplo yed Grad-CAM [22]. As sho wn in Figure 5, the heatmaps o v erlaid on four distinct ECG leads (I, II, V1, V5) visualize the re gions the model deems most important for its prediction. The high-intensity areas (yello w) consistently align with the QRS comple x es across all leads, suggesting it has learned to focus on v entricular depolarization the se gment of the ECG most af fected by L VH. (a) (b) Figure 4. Model performance analysis: (a) R OC curv e comparison and (b) training and v alidation loss curv es R WT -Net: A hybrid ResNet-wavelet-tr ansformer for early detection of left ventricular (Hoang Huu T o Nguyen) Evaluation Warning : The document was created with Spire.PDF for Python.
148 ❒ ISSN: 1693-6930 Figure 5. Model interpretability using Grad-CAM for a representati v e L VH case (ID: 3289), sho wing attention on QRS comple x es 3.4. Comparison with state-of-the-art T able 4 pro vides a detailed comparison of R WT -Net’ s performance ag ainst pre viously published re- sults for L VH det ection. While direct comparisons are challenging due to dif ferences in datasets, v alidation protocols, and patient populations, our proposed model demonstrates a superior performance prole, particu- larly in A UC. R WT -Net’ s high sensiti vity and rob ust v alidation on a lar ge dataset position it as a state-of-the-art solution. The performance of traditional criteria lik e Sok olo w-L yon is included to highlight the signicant leap in diagnostic capability of fered by machine learning approaches. T able 4. Detailed comparison with state-of-the-art methods for L VH detection Study Model/criterion A UC Sensiti vity Specicity Accurac y R WT -Net (this w ork) ResNet-wa v elet- transf ormer 0.9868 ± 0.0020 0.8964 ± 0.0292 0.9629 ± 0.0109 0.9502 ± 0.0062 Liu et al . [5] AI-enabled model 0.89 Not reported (NR) NR NR Khurshid et al . [7] Deep learning (L V mass prediction) 0.653 (UKB), 0.621 (MGB) NR NR NR Zhao et al . [6] CNN-LSTM 0.62 (test1), 0.59 (test2) 65–72% 57–71% NR Garza-Salazar et al . [23] C5.0 decision tree NR 79.6% (train/test), 81.6% (v al) 53% (train/test), 69.3% (v al) 71.4% (train/test), 73.3% (v al) Su et al . [1] Sok olo w-L yon 0.54 NR 95% NR Casale et al . [24] Cornell NR 52% 93% NR Le vy et al . [25] Sok olo w-L yon NR 6.9% 98.8% NR 4. CONCLUSION This study introduced R WT -Net, a no v el h ybrid deep learning model that demonstrates superior per - formance i n detecting L VH from 12-lead ECGs. By systematically inte grating a ResNet encoder , a statistical w a v elet feature stream, and a transformer fusion module, our model achie v es state-of-the-art performance on the PTB-XL dataset. The rigorous patient-strat ied cross-v alidation protocol ensures that our reported results are rob ust and generalizable. The k e y nding from our ablation study is the clinical utility of the w a v elet features; while not al w ays impro ving aggre g ate met rics lik e F1-score, the y signicantly boost the model’ s sensiti vity , which is paramount for an ef fecti v e early screening tool. Future w ork should focus on v alidating R WT -Net on di v erse, multi-center e xternal datasets to conrm its real-w orld rob ustness and ensure its general- izability across dif ferent ethnic populations. FUNDING INFORMA TION Authors state no funding in v olv ed. TELK OMNIKA T elecommun Comput El Control, V ol. 24, No. 1, February 2026: 142–150 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control ❒ 149 A UTHOR CONTRIB UTIONS ST A TEMENT This journal uses the Contrib utor Roles T axonomy (CRediT) to recognize indi vidual author contrib u- tions, reduce authorship disputes, and f acilitate collaboration. Name of A uthor C M So V a F o I R D O E V i Su P Fu Hoang Huu T o Nguyen ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Phuong Huu Nghia Le ✓ ✓ ✓ ✓ ✓ ✓ ✓ Lam Mai ✓ ✓ ✓ ✓ ✓ Nguyen Pham Ho T rong ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ C : C onceptualization I : I n v estig ation V i : V i sualization M : M ethodology R : R esources Su : Su pervision So : So ftw are D : D ata Curation P : P roject Administration V a : V a lidation O : Writing - O riginal Draft Fu : Fu nding Acquisition F o : F o rmal Analysis E : Writing - Re vie w & E diting CONFLICT OF INTEREST ST A TEMENT Authors state no conict of interest. D A T A A V AILABILITY The data supporting the ndings of this study are openly a v ailable in the PTB-XL dataset reposi tory at http://doi.or g/10.1038/s41597-020-0495-6, reference number [15]. REFERENCES [1] F .-Y . Su et al. , “ A comparison of Cornell and Sok olo w-L yon electrocardiographic criteria for left v entricular h ypertroph y in a military male population in T aiw an: the Cardiorespiratory tness and HospItalization Ev ents in armed F orces study , ” Car dio vascular Dia gnosis and Ther apy (CDT) , v ol. 7, no. 3, pp. 244–251, 2017, doi: 10.21037/cdt.2017.01.16. [2] X. Sun, Y . Y in, Q. Y ang, and T . Huo, “ Articial intelligence in cardio v ascular diseases: diagnostic and therapeutic perspecti v es, ” Eur opean J ournal of Medical Resear c h , v ol. 28, no. 1, p. 242, 2023, doi: 10.1186/s40001-023-01065-y . [3] Z. W u and C. Guo, “Deep learning and electrocardiograph y: systematic re vie w of current techniques in cardio v ascular disease diagnosis and management, ” Bio Medical Engineering Online , v ol. 24, no. 1, p. 23, 2025, doi: 10.1186/s12938-025-01349-w . [4] X. Liu, H. W ang, Z. Li, and L. Qin, “Deep learning in ECG diagnosis : A re vie w , ” Knowledg e-Based Systems , v ol. 227, p. 107187, 2021, doi: 10.1016/j.knosys.2021.107187. [5] C.-M . Liu et al. , “ Articial intelligence-enabled model for early detection of left v entricular h ypertroph y and mortality prediction in young to middle-aged adults, ” Cir culation: Car dio vascular Quality and Outcomes , v ol. 15, no. 8, p. e008360, 2022, doi: 10.1161/CIRCOUTCOMES.121.008360. [6] X. Zhao et al. , “Deep learning assessment of left v entricular h ypertroph y ba sed on electrocardiogram, ” F r ontier s in Car dio vascular Medicine , v ol. 9, p. 952089, 2022, doi: 10.3389/fcvm.2022.952089. [7] S. Khurshid et al. , “Deep learning to predict cardiac magnetic resonance-deri v ed left v entricular mass and h ypertroph y from 12-lead ECGs, ” Cir culation: Car dio vascular Ima ging , v ol. 14, no. 6, p. e012281, 2021, doi: 10.1161/CIRCIMA GING.120.012281. [8] J.-T . Huang et al. , “Comparison of machine learning and con v entional criteria in detecting left v entricular h ypertroph y and prognosis with electrocardiograph y , ” Eur opean Heart J ournal - Digital Health , v ol. 6, no. 2, pp. 252–260, 2025, doi: 10.1093/ehjdh/ztaf003. [9] N. Siranart et al. , “Diagnostic accurac y of articial intelligence in detecting left v entricular h ypertroph y by electrocardiograph: a systematic re vie w and meta-analysis, ” Scientic Reports , v ol. 14, no. 1, p. 15882, 2024, doi: 10.1038/s41598-024-66247-y . [10] K. C. Siontis, P . A. Nose w orth y , Z. I. Attia, and P . A. Friedman, “ Articial intelligence-enhanced electrocardiograph y in cardio v as- cular disease management, ” Nat. Re v . Car diol. , v ol. 18, no. 7, pp. 465–478, 2021, doi: 10.1038/s41569-020-00503-2. [11] S. L. Oh, E. Y . K. Ng, R. S. T an, and U. R. Acharya, “ Automated diagnosis of arrh ythmia using combination of CNN and LSTM techniques with v ariable length heart beats, ” Computer s in Biolo gy and Medicine , v ol. 99, pp. 165–175, 2018, doi: 10.1016/j.compbiomed.2018.06.002. [12] P . Madan, V . Singh, D. P . Singh, M. Diw akar , B. P ant, and A. Kishor , “ A h ybrid deep learning approach for ECG-based arrh ythmia classication, ” Bioengineering , v ol. 9, no. 4, p. 152, 2022, doi: 10.3390/bioengineering9040152. [13] R. Jothiramaling am, A. Jude, R. P atan, M. Ramachandran, J. H. Duraisamy , and A. H. Gandomi, “Machine learning-based left v entricular h ypertroph y detecti on using multi-lead ECG signal, ” Neur al Computing and Applications , v ol. 33, no. 9, pp. 4445–4455, 2021, doi: 10.1007/s00521-020-05238-2. [14] P . S. Addison, “W a v elet transforms and the ECG: a re vie w , ” Physiolo gical Measur ement , v ol. 26, no. 5, pp. R155-R199, 2005, doi: 10.1088/0967-3334/26/5/R01. [15] P . W agner et al .“PTB-XL, a lar ge publicly a v ailable el ectrocardiograph y dataset, ” Scientic Data , v ol. 7, no. 1, p. 154, 2020, doi: 10.1038/s41597-020-0495-6. [16] N. Strodthof f, P . W agner , T . Schaef fter , and W . Samek, “Deep learning for ECG analysis: Benchmarks and insights from PTB-XL, ” IEEE J ournal of Biomedical and Health Informatics , v ol. 25, no. 11, pp. 4119–4128, 2021, doi: 10.1109/JBHI.2020.3022989. R WT -Net: A hybrid ResNet-wavelet-tr ansformer for early detection of left ventricular (Hoang Huu T o Nguyen) Evaluation Warning : The document was created with Spire.PDF for Python.
150 ❒ ISSN: 1693-6930 [17] N. Strodthof f et al. , “PTB-XL+: A comprehensi v e electrocardiographic feature dataset, ” Scientic Data , v ol. 10, no. 1, p. 310, 2023, doi: 10.1038/s41597-023-02153-8. [18] T . Chen and C. Guestrin, “XGBoost: A scalable tree boosting system, ” in Pr oceedings of the 22nd A CM SIGKDD International Confer ence on Knowledg e Disco very and Data Mining , 2016, pp. 785–794, doi: 10.1145/2939672.2939785. [19] I. Loshchilo v and F . Hutter , “Decoupled weight decay re gularization, ” in International Confer ence on Learning Repr esentations (ICLR) , 2019. [20] F . Pedre gosa et al. , “Scikit-learn: Machine learning in Python, ” J ournal of Mac hine Learning Resear c h , v ol. 12, pp. 2825–2830, 2011. [21] C. C. Y ang, “Explainable articial intelligence for predicti v e modeling in healthcare, ” J ournal of Healthcar e Informatics Resear c h , v ol. 6, no. 2, pp. 228–239, 2022, doi: 10.1007/s41666-022-00114-1. [22] R. R. Selv araju, m. Cogswell, A. Das, R. V edantam, D. P arikh, and D. Batra, “Grad-CAM: V isual e xplanations from deep netw orks via gradient-based localization, ” arXiv , 2016, doi: 10.48550/arXi v .1610.02391. [23] F . l. Garza-Salazar , M. E. Romero-Ibar guengoitia, E. A. Rodriguez-D ´ ıaz, J. R. Azpiri-L ´ opez, and A. Gonz ´ alez-Cant ´ u, “Impro v ement of electrocardiographic diagnostic accurac y of left v entricular h ypertroph y using a machine learning approach, ” PLoS One , v ol. 15, no. 5, p. e0232657, 2020, doi: 10.1371/journal.pone.0232657. [24] P . N. Casale et al. “Electrocardiographic detection of left v entricular h ypertroph y: de v elopment and prospecti v e v alidation of impro v ed criteria, ” J ournal of the American Colle g e of Car diolo gy , v ol. 4, no. 1, pp. 202–210, 1984, doi: 10.1016/s0735- 1097(85)80115-7. [25] D. Le vy , S. B. Labib, K. M. Anderson, J. C. Christiansen, W . B. Kannel, and W . P . Castelli, “Determinants of sensiti vity and specicity of electrocardiographic criteria for left v entricular h ypertroph y , ” Cir culation , v ol. 81, no. 3, pp. 815–820, 1990, doi: 10.1161/01.CIR.81.3.815. BIOGRAPHIES OF A UTHORS Hoang Huu T o Nguy en is an emer ging researcher in the eld of articial intelligence, currently completing his studies at The Uni v ersity of Danang - V ietnam-K orea Uni v ersity of Informa- tion and Communication T echnology (VKU), V ietnam. His research is centered on the intersection of deep learning and biotechnology , with a specic focus on de v eloping no v el algorithms for biomedical time-series analysis . As the rst author of this study , he led t he architectural design and implementa- tion of the proposed R WT -Net frame w ork. He can be contacted at email: hoangnht.24ai@vku.udn.vn. Phuong Huu Nghia Le is a st udent at San Jacinto Colle ge, T e xas , where he is pursuing a de gree as a Biomedi cal Clinical Equipment T echnician. His technical studies focus on the princi- ples of medical de vice technology . F or this research, he contrib uted to the data curation, literature in v estig ation, and aspects of the softw are v alidation. His interests lie in the practical application of technology in clinical en vironments. He can be contacted at email: Le.P203416@stu.sanjac.edu. Lam Mai is a lecturer at The Uni v ersity of Danang - V ietnam-K orea Uni v ersity of In- formation and Communication T echnology (VKU), V ietnam. He recei v ed his Master’ s de gree from National Central Uni v ersity , T aiw an. His research interests include machine learning and pattern recognition. He can be contacted at email: mlam@vku.udn.vn. Nguy en Pham Ho T r ong is a lecturer in Information T echnology at FPT Uni v ersity , Da Nang, V i etnam. He earned his Mast er’ s de gree f rom Na tional Central Uni v ersity , T aiw an. His academic interests lie in inte grating theory with real-w orld data applications to support data-dri v en decision-making and applied machine learning. He can be contacted at email: nguyenpht@fe.edu.vn. TELK OMNIKA T elecommun Comput El Control, V ol. 24, No. 1, February 2026: 142–150 Evaluation Warning : The document was created with Spire.PDF for Python.