Computer Science and Mathematics

Abstract: This article examines issues related to determining the resolution in multichannel post-detector nonparametric signal detectors. We define resolution as the ability to distinguish closely spaced targets. Typically, theoretical studies examine the effectiveness of detecting a single useful signal in one of the Cell Under Test consistent with its informative parameters. However, in real systems, the number of useful signals can be large. Consequently, determining the ability to detect multiple signals is an extremely important task. This article discusses a method for determining resolution based on an analysis of the uncertainty function for Wilcoxon’s rank statistics.

Abstract: Although actigraphy is widespread across disciplines to quantify human locomotor activity, its methodology remains unstandardized: the measurement parameters used to record acceleration and the methods used to compress it into epoch-level activity signals vary substantially across manufacturers. While we previously characterized the impact of the latter by emulating device-specific activity determination methods, the effects of acceleration sampling frequency, bit resolution, and full-scale range remain unclear. To systematically investigate these effects, we had to characterize the correlation-based relationships among activity determination methods during sleep and wakefulness to distinguish differences arising from their distinct motion characteristics from those introduced by reducing acceleration data quality. Then, by modifying measurement parameters to emulate different hardware settings, we found that parameter alterations had no effect over multiday and wakeful timeframes. Even when focusing on sleep periods, modifying the measurement parameters affected only a few activity determination methods, with the extent of these effects also influenced by the post-calibration of acceleration data. Overall, how activity is derived from acceleration has a substantially greater influence on the resulting activity signals than the measurement parameters used to record the underlying acceleration data, although high-fidelity acceleration collection remains particularly important for calibration-dependent activity determination methods in sleep-focused analyses.

Abstract: The Minimum Variance Distortionless Response (MVDR) beamformer offers optimal spatial selectivity but suffers from severe numerical instability in non-stationary acoustic environments. Target encroachment, steering vector mismatch, and finite-snapshot variance cause the spatial covariance matrix (SCM) to become acutely ill-conditioned. As the matrix approaches singularity, the norm of the inverted SCM diverges, leading to unbounded filter weights and a catastrophic collapse in White Noise Gain (WNG). While trace-relative diagonal loading mitigates this instability, utilizing a fixed regularization parameter forces an intractable compromise between spatial selectivity and numerical robustness. This paper proposes a spatially informed, time-frequency adaptive regularization framework for MVDR beamformers. We derive an orthogonal subspace leakage metric that quantifies the projection error between the assumed target steering vector and the dominant spatial subspace of the recursively estimated mixture SCM. This metric serves as a computationally efficient, instantaneous proxy for spatial difficulty. By mapping this leakage through a continuous logistic sigmoid clutch bounded by an empirically validated, array-scaled transition point (1/M), the proposed method continuously adapts the trace-relative loading. Quantitative evaluations across varying array geometries demonstrate that this continuous mapping reduces frame-to-frame filter weight jitter by an order of magnitude compared to binary thresholding. The proposed architecture autonomously navigates the feasible performance frontier, bounding WNG collapse during spatial degradation while preserving maximum selectivity in favorable geometries, all without requiring environment-specific threshold tuning or iterative convex solvers.

Abstract: Freezing of gait (FoG) is one of the most debilitating episodic motor symptoms of Parkinson’s disease (PD). FoG detection from wrist worn inertial signals poses challenging problems such as extreme data class imbalance (only 12–18% of signal windows are FoG), low frequency postural instabilities alongside diagnostically crucial 3–8Hz rhythmic tremors, and the need for subject independent generalisation. To address these, this paper presents WaveFoG, a novel wavelet gated Transformer architecture that combines multi scale representations from discrete wavelet transform (DWT) with a dual branch encoder. A 1D CNN captures local gait texture, and a Transformer captures global temporal context. They are fused via a sigmoid gating fusion layer that modulates the encoder output based on the DWT sub-band descriptors. Trained with focal loss and evaluated on the public Kaggle TLVMC FoG Prediction dataset under subject independent grouped ten fold cross validation (SI-CV), WaveFoG achieves a mean F1-score of 0.875 ± 0.017 and AUPRC of 0.833 ± 0.020, exceeding single branch baselines (1D CNN,Bi-LSTM,andvanilla Transformer) by upto5.4pponF1. Temporalsaliency maps highlight temporal patterns that are consistent with reported characteristics of FoG onset. The source code of this project is available at: https://github.com/Shihabul-Shuvo/WaveFoG.

Abstract: Chirp coding and modulation is widely recognised as an effective approach for transmitting information through noisy and dynamically varying communication channels. However, whether chirp signalling is uniquely optimal relative to alternative modulation schemes under comparable channel and power constraints remains an important theoretical question. This paper develops a mathematical framework to investigate the optimality of chirp-based signalling in noise-dominated environments. Drawing on information theory, functional analysis, and time–frequency analysis, we examine the performance of chirp modulation with respect to information-theoretic efficiency, robustness to Doppler-induced frequency shifts, and resilience to multi-path and multiple-scattering propagation. Analytical comparisons with conventional modulation schemes are used to establish the conditions under which chirp signalling provides superior signal distinguishability, noise resilience, and propagation robustness. The analysis indicates that chirp-based waveforms possess a distinctive combination of time–frequency characteristics that enables them to maintain reliable signal separation under challenging channel conditions. Under the assumptions and channel models considered, the results suggest that chirp modulation can achieve simultaneous optimality across multiple performance dimensions, providing a stronger theoretical basis for its use than conventional modulation approaches. The results therefore support the proposition that chirp modulation provides a fundamentally advantageous, and under the specified conditions uniquely optimal, framework for reliable digital communication in noise-dominated environments.

Abstract: Sound is inherently ambiguous, which makes certainty critical in danger recognition, where false alarms and missed detection both carry real consequences. The definition of danger adopted here extends beyond ISO 7731:2003, which covers only intentionally designed alarm signals, to naturally occurring dangerous events such as glass break-ing, gunshots, and human distress sounds, following the acoustic surveillance litera-ture. Existing systems largely treat danger as multiclass classification without seman-tic reasoning about context, or rely on one-shot audio-to-word predictions that do not describe the scene. We instead treat danger as an outcome of the full acoustic scene, requiring semantically rich descriptions. The proposed framework trains a CNN for binary danger/no-danger classification on ESC-50, UrbanSound8K, and AudioSet, and in parallel extracts captions with the pretrained CoNeTTE model, scores them through a custom danger lexicon with SBERT similarity, and fine-tunes a DistilBERT classifier on the resulting labeled captions. The two classifiers are combined through a rule-based uncertainty fusion strategy, in which an inexpensive first stage decision re-solves clear-cut cases and the semantic classifier is reserved for ambiguous ones, re-flecting the low-cost deployment demands of practical systems. On UrbanSound8K, fusion achieves a danger class recall of 0.82 and F1-score of 0.67, improving over the standalone CNN (recall 0.70) while preserving a transparent, human readable decision pipeline. We further evaluate the rule on two external, unseen datasets (XD-Violence and DESED) from different acoustic domains to assess generalization.

Abstract: Automatic characterization of adventitious respiratory sounds is essential for developing computer aided systems that support respiratory disease diagnosis. Commonly used acoustic descriptors, such as Mel frequency cepstral coefficients (MFCCs), primarily characterize spectral information and may not fully capture the nonlinear temporal dynamics of respiratory signals. In this work, we investigate the discriminative capability of Generalized Weighted Ordinal Pattern (GWOP) descriptors for representing adventitious respiratory sounds. Using respiratory recordings from the publicly available SPRSound pediatric database, GWOP descriptors were computed over multiple embedding dimensions, weighting factors, and time delays. An XGBoost classifier was employed exclusively as a supervised validation model to assess the discriminative information contained in the proposed feature representation, while the XGBoost Gain criterion was used to analyze the contribution of individual descriptor components. The proposed representation achieved stable performance across cross validation folds, outperformed a conventional MFCC representation under the same evaluation protocol, and demonstrated that a reduced subset of Gain selected features preserved, and slightly improved, the classification performance while substantially reducing the dimensionality of the original feature space. These findings indicate that GWOP provides a lighter yet more informative and interpretable representation of adventitious respiratory sounds, capturing nonlinear ordinal dynamics that complement conventional spectral descriptors and offering a promising alternative for respiratory sound characterization.

Abstract: Sixth-generation (6G) wireless systems target ubiquitous connectivity by integrating terrestrial (TNs) and non-terrestrial networks (NTN) cooperation, including low-Earth orbit (LEO) satellites, high-altitude platform stations (HAPS), and unmanned aerial vehicles (UAVs). In such a vertically stratified architecture characterized by massive multi connectivity, the selection of the most suitable tier or tier-combination for each mobile user is a complicated task, as the decision depends jointly on the propagation environment, user requirements, and the instantaneous characteristics of the topology. This paper proposes an altitude-aware Deep Learning (DL) framework that casts multi-tier network selection as a seven-class classification problem spanning standalone TN, LEO, HAPS, and UAV access as well as their TN-assisted multi-connectivity combinations. A fourteen-feature physics-based dataset is generated through extensive simulations based on standardized channel and geometry models, including 3GPP TR 38.901 terrestrial pathloss and line-of-sight probability, satellite-constellation elevation angles, and air-to-ground link geometry for HAPS and UAV. A deep neural network (DNN) is trained on the aforementioned dataset and assessed using stratified five-fold cross-validation. The proposed model achieves an overall classification accuracy of 88.1% with a macro-averaged F1-score of 0.88 and a single-sample inference latency of approximately 2 ms, well within typical 6G handover budgets. The results demonstrate that requirement- and geometry-aware tier selection can be performed accurately, enabling real-time decision-making in ambiguous multi-connectivity configurations for 6G networks.

Abstract: Automatic speech recognition (ASR) evaluation for Garrusi, a Kurdish variety written in a Latin-based orthography, is complicated when an ASR system outputs Arabic script because direct scoring can confuse recognition errors with writing-system differences. This study evaluates MMS-1B-all with the Central Kurdish adapter on 1,722 speech segments from five Garrusi speakers (9,763 reference tokens; 117.9 minutes) without adaptation. A common-reference scoring design was used: the reference was folded once and kept fixed, while the hypothesis representation was changed. The unmodified Arabic-script output produced 111.70% WER and 100.92% CER. Transliteration to Latin reduced these to 102.36% and 57.89%, while additional orthographic folding reduced them to 97.45% and 51.13%. Thus, the full transformation reduced measured WER by 14.25 percentage points and CER by 49.79 points, but substantial error remained. A round-trip control showed that part of the remaining error comes from the transliteration and folding process itself, with 98.5% of its substitutions linked to two orthographic patterns. A separate comparison with a Southern Kurdish fine-tuned system showed a 10.69-point WER difference, which fell to 2.13 points when short i was treated as non-contrastive. These results show that orthographic representation can strongly affect ASR evaluation and that recognition performance should be assessed separately from differences introduced by the scoring pipeline.

Abstract: Intelligent systems deployed in smart cities, smart grids, environmental monitoring, and other data-driven applications increasingly depend on reliable multivariate time-series forecasting. Recent deep forecasting models have achieved strong performance on various benchmarks, but their final predictions are often generated through a fixed forecast-generation process. This one-size-fits-all approach may be suboptimal because input windows exhibit different values, trends, and periodic patterns. We propose gap-aware residual selection (GARS), a plug-in refinement module for time-series forecasting that improves the prediction ability of an existing forecasting model rather than replacing it. GARS constructs deterministic reference forecasts from the observed input window and uses the differences between these forecasts and the initial forecast of the model as gap-aware residual information. Using only the observed input window and initial forecast, GARS constructs three component forecasts: the initial forecast, a gap-aware residual component that adjusts the initial forecast using these reference differences, and a direct residual component learned from the input window. The final forecast is obtained by assigning soft weights to the components across different segments of the forecast horizon rather than using fixed combinations. Future target values are used exclusively for training and evaluation and not for constructing reference forecasts, residuals, or soft weights. Experiments on five multivariate benchmark datasets related to energy, environment, electricity, traffic, and finance, using four representative forecasting models, show that GARS consistently improves average forecasting performance, reducing the mean squared error by 7.95% on average compared with the original models.

Abstract: Encoding a one-dimensional time series as a two-dimensional image and classifying it with a convolutional neural network is a popular and accurate paradigm, yet the standard encoders, Gramian Angular Fields, Markov Transition Fields, recurrence plots, spectrograms, scalograms, and polar or spiral embeddings, are validated almost exclusively on downstream accuracy. A second, practically essential property is almost never measured: whether a human can look at the produced image and tell the classes apart. We propose the Generalized Iterative Polar Transform (GIPT), a readable time-series imaging framework, and present a significant and novel study that takes a polar-spiral encoder reporting strong benchmark numbers yet producing humanly unreadable images, and asks whether classifier accuracy and visual legibility can be achieved together. We make five contributions. First, we give an explicit geometric diagnosis proving why the existing encoder is opaque: the time term dominates the radius and collapses the angle, so the trajectory occupies a vanishing wedge and the number of spiral turns is provably inert. Second, we derive GIPT, a principled readable encoding that swaps the polar axes, renders a filled silhouette, and uses an amplitude-aware normalization, and we prove a lobe-counting property that renders a frequency difference countable by eye. Third, we isolate a genuine readability-accuracy trade-off and show it is governed entirely by the normalization. Fourth, we show a striking and useful consequence: because GIPT enforces cross-sample consistency, a parameter-free one-shot nearest-neighbor classifier on the encoded images matches or exceeds the convolutional network at a fraction of the cost, eliminating the need for a trained model. Fifth, we propose a paired evaluation protocol that couples a stabilized multi-seed accuracy benchmark with a vision zero-shot legibility test and a human test-set study, turning legibility into a number on the same images. Across a broad suite of twenty-eight univariate UCR archive datasets the GIPT multi-channel encoder is competitive with the opaque baseline (76.0% vs 73.8% CNN accuracy), the training-free one-shot classifier on GIPT images reaches 83.1%, on par with the classical DTW nearest-neighbor standard while running about 1,200 times faster than training the network, and the human-facing rendering is genuinely eye-classifiable where the class difference is morphological, a property we verify with ablation, robustness, and statistical significance analyses. We conclude that readability and accuracy are not in conflict; they are reconciled by rendering the human-facing and network-facing images under different normalizations.

Abstract: Surface electromyography (sEMG) signals acquired with low-cost sensors tend to exhibit variable degradation that can compromise the reliability of myoelectric control systems outside controlled conditions. This work presents an adaptive processing pipeline for sEMG signals composed of three chained stages: a motor-intent classifier based on dilated temporal convolutional networks, whose function is to determine whether a signal window contains muscle activity associated with a voluntary gesture; a dual-output quality assessor that estimates a continuous score and a binary acceptability label, aimed at deciding whether the signal can be used directly or requires intervention; and a convolutional autoencoder that recovers the morphology of degraded windows before they are used in prosthetic control. The decision policy for reconstruction operates on two independent thresholds and classifies each window into one of three states: signal discard, direct acceptance, or active restoration. The models within the proposed architecture are trained on the public NinaPro DB1, DB3, and DB10 datasets and validated without recalibration on signals recorded from two participants with transtibial amputation over four weekly sessions. The results show that the pipeline correctly handles signal profiles with opposing characteristics. Inference latency remained below 5 ms at the 50th percentile across all scenarios, consistent with real-time operation. An integrated prototype using an Arduino UNO R4 WiFi was used to evaluate the results, and validate the feasibility of the system on physical hardware.

Abstract: Photon-counting computed tomography (PCCT) represents a detector-level transformation in CT imaging. Unlike conventional energy-integrating detectors, photon-counting detectors directly convert individual X-ray interactions into electrical pulses and classify them according to energy. This architecture enables electronic-noise rejection, smaller detector pixels, improved geometric dose efficiency, and intrinsic spectral acquisition. However, the images available to radiologists are not produced directly by the detector; energy-resolved photon counts must first undergo calibration, correction, projection formation, reconstruction, and material decomposition. This narrative review provides an educational framework linking X-ray attenuation physics, detector materials and architectures, energy thresholds, and detector nonidealities to the resulting PCCT images. It describes conventional polyenergetic and ultra-high-resolution images, virtual monoenergetic imaging, iodine maps, virtual non-contrast imaging, calcium and bone subtraction, virtual non-calcium imaging, effective atomic number maps, electron-density maps, and emerging K-edge techniques. Particular emphasis is placed on the clinical purpose and limitations of each reconstruction, including noise, artifacts, partial-volume effects, misregistration, incomplete subtraction, calibration dependence, and limited cross-platform comparability. Practical considerations for protocol design, image selection, interpretation workflow, and spectral-data archiving are also discussed. Understanding the pathway from photon detection to image formation is essential for selecting the appropriate reconstruction, avoiding misinterpretation, and integrating PCCT effectively into clinical radiology.

Abstract: This paper proposes a switching mode beamforming (SMVB) technique for a two-microphone array setup in a multi-source acoustic environment with one dominant target and two interferers. The proposed method adaptively switches between a robust minimum variance beamformer (RMVB) and a linearly constrained minimum variance (LCMV) beamformer based on the eigenvalue characteristics of the spatial covariance matrix. The eigenvalue distribution is used to infer the signal environment and select the appropriate beamforming strategy. Simulation results show that the proposed SMVB method outperforms conventional MVDR, RMVB, and LCMV beamformers in terms of interference suppression and target signal preservation, achieving improved output signal-to-interference-plus-noise ratio (SINR) and overall signal quality.

Abstract: Always-on keyword spotting (KWS) on microcontrollers must reconcile high recognition accuracy with severe constraints on memory, computation, and energy. Most small-footprint KWS research targets English, and its conclusions remain largely unverified for typologically different, consonant-rich languages such as Polish. In this work, we present a complete, hardware-validated KWS solution for Polish on the RP2350 microcontroller (Raspberry Pi Pico 2). We design three compact architectures from the convolutional (CNN), convolutional-recurrent (CRNN), and depthwise-separable (DS-CNN) families and benchmark them against BC-ResNet, a state-of-the-art reference, on a 25-keyword Polish vocabulary using MFCC features. All models are evaluated in full precision (Float32) and after 8-bit integer (INT8) post-training quantization, with inference latency measured directly on the target hardware. BC-ResNet attains the highest full-precision accuracy (97.81%) but is the most fragile under quantization, whereas the proposed CRNN becomes the most accurate quantized model (94.28%), the DS-CNN the smallest (46.15 KB), and the CNN the fastest (107.5 ms); all quantized models meet a one-second real-time budget. We further show that the accuracy ranking inverts after quantization, that memory savings are highly architecture-dependent, and that phonetically similar Polish words are the dominant source of error. These results offer practical guidance for deploying small-footprint KWS in Polish and other under-represented languages.

Abstract: Fine-grained and open-set fault diagnosis of analog circuits poses two challenges that generic deep classifiers fail to address. First, under a closed-set assumption, models assign unseen fault types to known classes with unwarranted confidence, lacking any mechanism to reject out-of-distribution inputs. Second, component manufacturing tolerances induce structured intra-class variation that causes the frequency responses of distinct fault modes to overlap, severely degrading fine-grained discrimination. This paper proposes a Tolerance-Aware Contrastive Siamese Network (TCSN), a metric-learning framework that constructs a discriminative embedding space jointly addressing both issues. The core contribution is the Tolerance-Aware Contrastive Loss (TACL), comprising a tolerance term \( (\mathcal{L}_{\mathrm{tol}}) \) that suppresses tolerance-induced intra-class scatter across Monte Carlo realizations, and a fine-grained term \( (\mathcal{L}_{\mathrm{fg}}) \) that enforces separation of near-identical overlapping classes. For open-set rejection, an energy-based scoring mechanism maps prototype distances into in-/out-of-distribution scores, identifying unknown faults without retraining. The framework is validated on a Sallen-Key second-order Butterworth low-pass filter benchmark (10 known and 2 unknown fault classes). It attains an open-set AUROC of 0.9309 and an FPR@95%TPR of 0.2500; for closed-set classification it reaches 90% accuracy (macro-F1: 0.88), and it retains 83.0% accuracy under one-shot conditions (n = 1). Under corrupted inputs it degrades gracefully, outperforming all baselines at 10 dB SNR. Ablations confirm the necessity of each component: removing \( \mathcal{L}_{\mathrm{tol}} \) drops the open-set AUROC by 0.3793, removing \( \mathcal{L}_{\mathrm{fg}} \) reduces fine-grained accuracy by 11 percentage points, and replacing energy scoring with distance thresholding lowers the AUROC by 0.2683.

Abstract: The digitized models of Chinese traditional musical instruments have fallen behind those of Western musical instruments, therefore they have restricted the preservation work and the research acquiring. This article deals with score-to-audio generation for the Chinese bamboo flute (dizi), it produces expressive audio from MIDI scores which has human performance detail. The current existing methods have two obstacles that they face. First, the large-scale paired training data (MIDI-audio) for traditional Chinese musical instruments is not much. Second, the already existing systems are not able to control the techniques that are specific to instruments—vibrato, pitch bends, ornaments—which the standard MIDI has no ability to encode. We put forward a expandable working flow that produces infinite artificial MIDI-audio matching groups through rule-based music score random arrangement, automatic DAW drawing output by using business sample storage banks, and systemized character pick-up. The information of technique is directly coded in the MIDI stream through the use of note numbers that lie outside the normal range, which thus proves to be more effective than the method of external conditioning. This system is constructed on the basis of MIDI-DDSP. Pre-training makes use of large-scale man-made data to obtain timbre consistency; The fine adjustment carries out the use of actual recording materials (URMP flute corpus) to realize expressive dynamic changes. Only the synthetic data can generate output that is stable but has no life. The real data which exist alone can bring about the problem of overfitting. This combined result can give realistic tone, which has controllable dynamic changes and playing skills.

Abstract: The accuracy of spectro-temporal features for Makhaarij al-Huroof and Sifaat distinguishes between the ten canonical Qiraát recitation styles of the Holy Quran. However, real-world room reverberations blur formant contours and corrupt inter-word energies, thus making Qiraat discrimination difficult. The current dereverberation methods were designed to work under ordinary speech conditions and are not capable of preserving phonetic qualities for domain-specific purposes. This paper introduces a four-step, phase-consistent signal processing approach prioritizing phonetic preservation over direct reverberation suppression. The four steps are: (1) adaptive noise floor attenuation; (2) soft voice activity detection using power-law boundary decay; (3) application-specific spectral contour adjustment from clean Quranic reference audio; and (4) Griffin-Lim algorithm-based phase correction. A total of 48 real-world room recordings were utilized for the evaluation of this approach based on Energy Ratio (ER), Spectral Contrast (SC), and Formant Clarity (FC) – measures specific to the Quran audio domain – alongside conventional speech quality metrics. The proposed approach yielded the highest scores in four out of seven metrics, namely ER (+19.58 dB), SC (+40.11), FC (+822.94), and PESQ (+1.251), while being superior to Spectral Subtraction, Wiener Filtering and WPE Dereverberation approaches. Moreover, the perceptual enhancement was verified in a synthetic controlled experiment where the proposed approach scored an improved PESQ metric (+2.495; SNR −1.874dB). The results illustrate the fact that an optimization for general purpose metrics does not necessarily ensure phonetic preservation required for specific classification.

Abstract: Interpersonal synchrony is a time-dependent coordination pattern in which interacting partners' body movements become temporally aligned. This study frames interpersonal synchrony as a human motion analysis problem and presents Synchrony Vision, an RGB-D sensor-based system for real-time monitoring and event-level analysis of interpersonal motion synchrony in free dialog. The system transforms Kinect-derived skeletal positions into joint acceleration signals, applies sensor-specific conditioning, detects movement peaks, and estimates event-level phase differences between two participants within a ±1.0 s window. The operator-facing interface supports live RGB-D monitoring, acceleration visualization, joint selection, millisecond-scale phase-difference histograms, four synchrony metrics (Frequency, Direction, Width, and Strength), and exportable acceleration, timestamp, peak-pairing, and summary artifacts. We evaluated the deployed pipeline on 25 Kinect-tracked dyads engaged in unconstrained conversation. Across 200.6 min comprising 245,835 frames, the system detected 2,681 synchrony events, exceeded circular-surrogate baselines, and preserved between-dyad ordering across reasonable parameter settings. Motion Energy Analysis-style cross-correlation on the same acceleration signals also confirmed above-chance synchrony but produced different dyad rankings. These findings show that RGB-D skeletal sensing can extend human motion analysis from individual movement capture to transparent, event-level quantification of interpersonal coordination.

Abstract: Accurate R-peak detection in electrocardiogram (ECG) signals is fundamental for cardiovascular analysis. However, most existing methods address differences in sampling frequency (fs) through signal resampling or transfer learning, which may alter the temporal definition of annotated events. In this study, we propose a fs consistent framework for ECG R-peak detection that avoids both resampling and retraining. The proposed method is based on low-sampling morphological learning combined with physiological temporal constraints (PTC). A lightweight classifier (Extreme Gradient Boosting) is trained on 128 Hz ECG data (MIT-BIH Normal Sinus Rhythm Database, XGB) to learn local morphological structures, and feature extraction is defined in milliseconds with time-normalized derivatives to ensure consistency across fs. The trained model is directly applied to higher- fs datasets (360 Hz, 500 Hz, and 1000 Hz) without modification. Final peak locations are determined through deterministic processing, including PTC and local snap processing. Experimental results demonstrated that the proposed method achieved stable detection performance across multiple sampling frequencies. When evaluated in a sample-wise manner, the proposed method achieved mean F1-scores of 0.885 on MIT-BIH Arrhythmia Database (360 Hz), 0.848 on Lobachevsky University Electrocardiography Database (LUDB, 500 Hz, sinus rhythm), 0.837 on LUDB (500 Hz, arrhythmia), and 0.953 on PTB Diagnostic ECG Database (1000 Hz), without any resampling or retraining. The integration of probabilistic candidate detection and deterministic temporal alignment enables consistent peak localization under cross-frequency conditions. These findings demonstrate that augmenting machine learning with deterministic decision mechanisms provides a principled framework for fs -consistent ECG peak detection.