1.1. Background

Cardiovascular diseases (CVDs) remain the leading cause of mortality worldwide, responsible for nearly 18 million deaths each year [1]. Early detection and continuous monitoring are essential to enable timely intervention, improve outcomes, and ease the burden on healthcare systems. While electrocardiography (ECG) is the clinical gold standard for cardiac monitoring, its multi-electrode configuration and the need for clinical supervision limit its suitability for pervasive, long-term use [2]. Photoplethysmography (PPG), on the other hand, is a non-invasive, low-cost optical method that measures changes in blood volume at the microvascular level. PPG is a prevalent method for conducting extensive, decentralized health assessments and preventative cardiac care because of its extensive application in smartwatches and activity bands [3]. PPG serves as a valuable tool for context, quality assurance, and assessing cardio-metabolic risk. PPG has transitioned from simple pulse oximetry to a versatile platform for disease screening and cuffless hemodynamic monitoring. Early spectral methods such as MUSIC-based analysis on PPG demonstrated discriminative capacity between healthy and athletic subjects and established practical preprocessing and peak-detection considerations [4]. More recently, feature-selection pipelines on PPG have identified coronary artery disease with clinically relevant accuracy, emphasizing interpretable features and pragmatic selection strategies that can be repurposed for quality-aware decision support in home monitoring [5]. Deep-learning approaches using PPG scalograms and PPG-NET have shown encouraging results for non-invasive hypertension monitoring and for blood-pressure stage classification with pre-trained models [6–7]. Complementary engineering studies comparing piezoelectric, photosensor, and earlobe capture clarify trade-offs in signal fidelity and robustness across form factors—evidence directly applicable to selecting a PPG sensor site that coexists with a breath-sampling mouthpiece [8].

1.2. Problem Statement

The PPG waveform encodes a wealth of physiological information about cardiac pulsation, vascular elasticity, and autonomic control [9]. Alterations in waveform morphology, rhythm, or amplitude can reveal diverse pathologies ranging from arrhythmias to heart failure. Prior studies have demonstrated the use of PPG in detecting specific conditions such as atrial fibrillation (AF) [10] and sleep-related disorders [11]. However, a clear gap persists in developing robust, unified models capable of discriminating across multiple cardiac anomalies from raw PPG signals. This is primarily due to subtle inter-class morphological variations that traditional signal processing or shallow machine learning models struggle to capture [12].

1.3. Justification of the Problem

Existing approaches for multi-class CVD detection often depend on labor-intensive, hand-crafted feature extraction—such as crest time, reflection index, or stiffness indices—followed by conventional classifiers like Support Vector Machines or Random Forests [5,13]. While such methods can identify coarse distinctions, they are limited in capturing fine temporal details. Meanwhile, recent deep learning models optimized for wearables emphasize computational efficiency, but at times sacrifice the representational depth needed to characterize complex cardiac conditions [14,15].

1.4. Bridging the Gap: Proposed Approach

A deeper hierarchical 1D Convolutional Neural Network (CNN), trained end-to-end on raw PPG signals, is hypothesized in this work that it can capture the multiscale feature representations needed to distinguish multiple cardiovascular conditions with high fidelity. Unlike lightweight architectures that compromise expressive power, our proposed design consists of four stacked convolutional blocks of increasing depth. This design enables the progressive extraction of features, including high-level pathological characteristics and low-level pulse dynamics, without the need for manual feature engineering.

1.5. Main Highlights and Contributions

This research presents many contributions; among the main are the following:

1. Innovative Architecture: A specialized deep 1D-CNN engineered to exploit the temporal and morphological complexity of PPG signals for effective multi-class differentiation.
2. End-to-End Learning: The model obviates the need for manual feature engineering by learning directly from preprocessed raw PPG signals.
3. Comprehensive Classification: Demonstrates strong performance in distinguishing six cardiac conditions, extending beyond the narrow binary focus common in prior work.
4. Robust Performance: Achieves state-of-the-art accuracy (93.48%), confirming the importance of deeper architecture for complex PPG-based diagnostics.

1.6. Structure of the Article

The remainder of this paper is organized as follows. Section 2 reviews related work on PPG-based cardiovascular disease detection. Section 3 describes the dataset, preprocessing pipeline, and the proposed CNN model. Section 4 presents experimental results and comparisons. Section 5 discusses methodology, clinical implications, and interpretability. Section 6 sheds light on main limitations and future work. Section 7 concludes with a summary of findings.

3.1. Data acquisition and preprocessing pipeline

High-quality, consistently processed PPG is essential for reliable learning. The dataset comprised 2,448 labeled PPG segments pooled from public corpora (e.g., clinical PPG-BP derivatives, MIMIC-style sources) and curated clinical collections, with cardiologist-confirmed labels for AF, HF, ACS, CVA, DVT, and Normal. Fixed 12-second windows (~12–20 beats at 60–100 bpm) were used to balance temporal context with stationarity, enabling simultaneous capture of rhythm irregularity (AF) and morphology-linked pathology (HF) without excessive label drift—an approach consistent with reviews of PPG deep learning and AF screening in real-world smartphone workflows [18,30,35–38]. Fig 1 shows the block diagram of the multiclass-PPG model.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Block diagram of the multiclass-PPG model.

https://doi.org/10.1371/journal.pone.0347840.g001

3.2. Dataset composition and clinical characteristics

The dataset analyzed in this study comprises 2,448 photoplethysmography (PPG) signal segments derived from 612 unique patients collected from a multi-source clinical pool. Data was obtained from a tertiary-care cardiovascular center and affiliated diagnostic laboratories between 2016 and 2024 under institutional approval. All diagnoses were confirmed by board-certified cardiologists based on clinical examination, imaging, laboratory findings, and established diagnostic criteria.

The distribution of unique patients across diagnostic classes was as follows:

• Atrial Fibrillation (AF): 94 patients
• Heart Failure (HF): 88 patients
• Acute Coronary Syndrome (ACS): 102 patients
• Cerebrovascular Accident (CVA): 96 patients
• Deep Vein Thrombosis (DVT): 87 patients
• Normal Sinus Rhythm (NSR): 145 patients

The mean age of the cohort was 57.4 ± 13.2 years, with 58% male and 42% female participants. Comorbidities such as hypertension and diabetes were recorded when available but were not used as predictive features.

Inclusion criteria consisted of:

1. Confirmed clinical diagnosis of one of the target conditions.
2. High-quality PPG recording of at least 12 seconds duration.
3. Age ≥ 18 years.

Exclusion criteria included:

1. Severe motion artifacts compromise signal morphology.
2. Incomplete clinical documentation.
3. Coexisting acute systemic instability.

The study protocol was approved by the Institutional Review Board of the participating healthcare institutions using the IRB/2016/215, IRB/2021/115, and IRB/2024/333, and all data were anonymized prior to analysis.

3.3. Patient-wise splitting and data leakage control

3.4. Proposed deep hierarchical convolutional neural network architecture

Robust multi-class PPG diagnosis benefits from hierarchically structured models that learn multi-scale temporal morphology—not solely ultra-lightweight AF-specific designs. The proposed 1D-CNN comprises four convolutional blocks (filters 32 → 64 → 128 → 256; kernels 13–15) with BN-ReLU-MaxPool in each block, followed by global average pooling and a regularized classifier head (dense-128 with L2 = 0.01, dropout = 0.5; softmax). The table you provided already enumerates roles and clinical advantages; below are literature-anchored design rationales.

3.4.1. Architecture overview.

The network comprises four sequential convolutional blocks with increasing filter depth, followed by a classification head. Table 1 summarizes the hierarchical CNN architecture for PPG analysis, detailing each layer’s specifications, functional role, and clinical advantages, highlighting how progressive feature extraction enhances robustness, interpretability, and multi-class cardiovascular classification accuracy. This hierarchy enables progressive learning from basic pulse structures to pathological morphologies:

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Structured Representation of the Deep CNN for PPG Signal Analysis.

https://doi.org/10.1371/journal.pone.0347840.t001

Distinguishing Characteristics:

1. 1. Hierarchical Morphology Learning:

Early layers capture pulse edges and local slopes; mid-layers integrate beat-to-beat timing; deeper layers synthesize pathological signatures (e.g., low-amplitude HF, irregular AF morphology). Reviews and recent PPG models for multi-class arrhythmia emphasize the advantage of multi-scale temporal feature extraction for generalization beyond binary tasks [30,45–47].

1. 2. Domain-Optimized Kernel Sizes:

Medium/large temporal kernels (≈65–75 ms at 13.33–15.4 Hz; proportionally scaled at lower Fs) align with sub-intervals of the systolic upstroke and notch region, improving efficiency over vision-style small kernels by embedding physiological priors—a trend echoed in large-kernel re-parameterization work for sensor time-series [48] and in arrhythmia-focused CNN/Transformer hybrids [45,49].

1. 3. Progressive Pooling and Aggressive Reduction:

Max-pooling (2,2,2,4) reduces sensitivity to small temporal shifts and suppresses residual noise while keeping discriminative envelopes—an approach consistent with computationally efficient PPG pipelines targeting on-device feasibility [15,50,51].

1. 4. Enhanced Regularization:

Batch normalization, global average pooling, and dropout jointly curb overfitting and improve out-of-distribution performance, a recurrent challenge in PPG deployment across sites/devices [30,37,38,41,52].

Comparative Positioning:

• vs. Lightweight DSCNNs richer multi-scale features enable six-class diagnosis rather than narrow AF screening [18,38,45].
• vs. CNN-LSTM/Transformer hybrids: comparable representational depth with fewer parameters and stable convergence; Transformers are compelling in very large datasets, but 1D-CNNs remain advantageous for edge efficiency and latency [45,46,49,51].
• vs. handcrafted ML: end-to-end learning surpasses the brittleness of feature-engineered pipelines while remaining interpretable through derivative-aligned saliency and notch/crest attribution maps [40,52].

3.5. Experimental setup and training protocol

Implementation used TensorFlow 2.8 on NVIDIA A100s. A stratified 80/10/10 split preserved class balance across train/val/test. Optimization used AdamW (β₁ = 0.9, β₂ = 0.999, ε = 1e-7, weight decay 0.01) with lr = 1e-4, halved on plateau (10-epoch patience); early stopping (patience = 20) curtailed overfitting. To address residual imbalance, class-weighted cross-entropy was used; label smoothing (ε = 0.05) stabilized calibration—both practices commonly adopted in multi-class PPG studies and reviews [30,45–47].

Evaluation metrics. Accuracy, macro-F1, Cohen’s κ (agreement beyond chance), MCC (balanced performance under skew), and Brier score (probability calibration) were reported. κ interpretation followed Landis–Koch [36]. The emphasis on beyond-accuracy metrics reflects best practice in recent reviews and clinical validations for AF and BP estimation using PPG [18,35,37,38].

4.1. Overall performance

On the held-out test set (n = 2,448 segments), the model achieved accuracy 93.48% (95% CI 92.2–94.6), macro-F1 = 0.9386 (95% CI 0.925–0.952), κ = 0.8968 (95% CI 0.882–0.911; “almost perfect” [36]), balanced accuracy 92.9% (95% CI 91.5–94.1), and Brier = 0.071 with well-behaved calibration curves. These metrics align with the performance envelope reported for recent multi-class arrhythmia and rhythm-disorder models on wearable PPG while extending to a broader diagnostic label set [45–47]. Sample run of the performance metrics of the proposed model is detailed in Table 2.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Comprehensive performance metrics of the proposed deep CNN model on the independent test set.

https://doi.org/10.1371/journal.pone.0347840.t002

The model achieved a remarkable overall accuracy of 93.48%, with the validation accuracy converging to an identical value, indicating a stable training process devoid of overfitting. The Cohen’s Kappa score of 0.8968 falls within the range of “almost perfect agreement” (0.81–1.00) according to Landis and Koch’s benchmark, signifying that the model’s predictions are 89.68% more accurate than those expected by random chance alone. In medical applications, when agreement exceeding chance is critical, this is an essential parameter.

4.1.2. Error analysis.

The normalized confusion matrix (Fig 2) indicated that:

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. A Sample of the unbalanced classes confusion matrix.

https://doi.org/10.1371/journal.pone.0347840.g002

• 6.4% of CVA cases were misclassified as ACS.
• 9.7% of DVT cases were misclassified as Normal.
• No AF or HF cases were misclassified, highlighting their distinct signal morphologies.

The above results mirror literature that AF/HF often manifest with pronounced PPG differences while ischemic/cerebrovascular conditions exhibit partially overlapping peripheral signatures [45,53,54]. Overlap indices (e.g., Jaccard ≈ 0.31 for ACS ↔ CVA) supported physiologic plausibility rather than random error. Because multiple segments were derived from each patient, strict patient-level partitioning was enforced to eliminate data leakage and prevent artificially inflated performance.

4.2. Comparative benchmarking against contemporary architectures

To provide a rigorous evaluation against established baselines, we conducted two forms of comparison.

4.2.2. Internal baseline re-implementation.

To ensure a controlled comparison, we implemented a lightweight depthwise-separable CNN (DS-CNN) baseline inspired by recent wearable-oriented architectures. The model was trained using the identical patient-wise split. The results are shown in Table 3.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Comparative Analysis of PPG-based Cardiac Disease Detection Methods.

https://doi.org/10.1371/journal.pone.0347840.t003

Download:

• PNG
  larger image
• TIFF
  original image

https://doi.org/10.1371/journal.pone.0347840.t004

Additionally, the proposed model was compared with hybrid architectures such as Res-BiANet and CNN-Transformer frameworks reported in recent PPG literature. The proposed model demonstrated a statistically significant improvement in macro-F1 (p < 0.01, McNemar test), indicating that the additional representational depth enhances multi-class discrimination without excessive computational burden.

Inference latency remained below 5 ms per segment on consumer-grade hardware for both architectures.

4.3. Benchmarking against state-of-the-art

Wearable multi-class rhythm disorder detection: Recent studies using smartwatch/wearable PPG report strong but task-limited performance (AF-centric or rhythm-subset) in real-life settings [35,45–47,55]. Against that backdrop, the present six-class framework maintains comparable AF performance while adding HF/ACS/CVA/DVT classes—an expansion encouraged by contemporary reviews and hybrid CNN/Transformer advances [30,45,46,49,56].

• HF screening from PPG: Machine-learning approaches for HFpEF/HFrEF using PPG have shown feasibility [53,57]; the perfect HF detection here (within a multi-class setting) is consistent with morphology-sensitive pipelines when preprocessing/normalization and hierarchical temporal features are carefully designed [30,53,57].
• BP/vascular correlates: Spectro-temporal features and derivative indices are established surrogates for vascular load and BP estimation [13,37]; the observed Normal vs disease separability and ACS/CVA overlap are consistent with these biomechanical dependencies.

4.4. Robustness and clinical utility

Robustness testing under simulated noise demonstrated minimal performance degradation:

Under simulated noise, Gaussian σ = 0.05 reduced accuracy by 1.7%; motion perturbations by 2.1%, aligning with gains reported by artifact-aware and tiny edge models [15,44]. Decision-curve analysis (DCA) showed net benefit over treat-all/none strategies across clinically relevant thresholds, supporting translational utility for triage and alerts [18,38,58].

5.1. Confronting the incongruity: Balancing power and efficiency in representation

A persistent debate in PPG analytics concerns the trade-off between representational power and deployability on constrained devices. Ultra-lightweight architectures (TinyML, DSCNN) prioritize latency and footprint but typically target binary AF detection where rhythm irregularity dominates and morphological richness is not fully exploited [15,18,38]. Conversely, hybrid CNN–LSTM/Transformer stacks can capture longer-range dependencies but often carry large parameter counts and training instability, complicating clinical translation and edge inference [45,46,49,56]. The present hierarchy resolves this tension by using domain-optimized, medium/large kernels (13–15) and progressive pooling to encode multi-scale morphology (upstroke, notch, diastolic decay) while keeping inference costs compatible with smartphones and mid-tier edge devices, in line with sensor-series design evidence on large kernels and edge-aware practices [15,48]. The block-wise progression—from local edges to composite signatures—mirrors physiological reasoning and helps separate high-contrast phenotypes (AF, HF) from partially overlapping ischemic/cerebrovascular patterns [9,13,40,45,46]. Such architecture choices are aligned with recent reviews calling for models that balance accuracy with on-device feasibility and robustness under ambulatory conditions [15,30,51,56].

5.2. Eliminating the need for singular class prioritization and hand-crafted features

Historically, PPG diagnostics relied on handcrafted indices (timings, amplitude ratios, derivative peaks), which, while interpretable, restrict learning to preselected descriptors and may underperform when pathology manifests in subtle morphology beyond those features [13,36,59]. End-to-end learning over raw/time–frequency PPG offers a more complete representation, with strong evidence for superior performance in ambulatory heart-rate estimation and AF detection where motion and illumination induce complex distortions [15,18,38]. The current framework demonstrates flawless AF and HF discrimination and high sensitivity for ACS within a six-class setting, consistent with reports that morphology-aware CNNs (and measured upstroke/derivative dynamics) capture hemodynamic change and autonomic tone better than fixed feature sets [13,42,45,47]. Moving beyond single-class screening is clinically meaningful: AF screening alone, even at population scale, addresses only one pathway of cardiovascular risk [38,60]. Expanding to HF, ACS, and CVA aligns with reviews urging broader CVD stratification from PPG and with consensus pathways that embed wearables into triage and follow-up rather than siloed detection [37,38,56,58].

5.3. Enhanced generalization through robust preprocessing and regularization

Generalization gaps remain the Achilles’ heel for medical AI: performance degrades with new devices, skin tones, sampling sites, and contexts [30,37,45]. The dual-stage normalization (per-segment Z-score then min–max) reduces amplitude dispersion and stabilizes gradients across device/site variance; physics-aware denoising preserves notch/crest morphology while preventing over-smoothing that erases diagnostically relevant reflections [9,13,40,43]. Architectural regularization (batch normalization, global average pooling, dropout) curbs overfitting and supports calibrated probabilities—critical for downstream clinical decision tools and net-benefit analysis [61]. Observed confusions (e.g., ACS ↔ CVA) are physiologically plausible, reflecting shared peripheral hemodynamics rather than random error, a behavior echoed in real-world wearable validations and hybrid models that report overlap among ischemic/cerebrovascular phenotypes on PPG [35,45,47,53,55,56]. Artifact controls (PSD/DC-drift checks; motion-artifact screening) accord with emerging best practice for remote PPG acquisition and tiny edge models that mitigate illumination/motion corruption during inference [15,43,44].

5.4. Comparative Analysis and Niche Positioning

The following comparative table positions our contribution against prior influential studies, highlighting how it systematically addresses their weaknesses:

The framework is positioned against representative studies to highlight scope, efficiency, and robustness:

Pereira et al. (2020)—Handcrafted features + ML; 2 classes (AF/Normal); strength: interpretability; limitation: narrow rhythm scope → advance: end-to-end morphology across 6 classes [18].

lightweight AF—RPS/CNN; 2 classes; strength: edge efficiency; limitation: morphology underfit → advance: deeper multi-scale representation for complex pathology while maintaining edge feasibility [15,18,19].

Hybrid CNN–LSTM/Transformer—multi-scale temporal modeling; strength: accuracy; limitation: parameter/latency burden → advance: balanced depth-efficiency with stable training and on-device latency [45,46,49,56].

Res-BiANet multi-class PPG—DCNN for arrhythmias; strength: multi-class precedent; limitation: wearable-deployment details limited → advance: six-class scope with artifact/SQI controls and calibration reporting [45–47].

BP/vascular surrogates (spectro-temporal)—link morphology to vascular state; strength: physiology-anchored features; limitation: not disease-specific → advance: integrates morphology into disease classification with clinical metrics beyond accuracy [13,37,59].

XAI readiness—saliency/attribution on PPG matrices/scalograms; strength: transparency; limitation: cost → advance: calibrated base model compatible with Grad-CAM/LRP/SHAP for clinical reasoning [46,52].

By addressing hand-engineered fragility, AF-only scope, and heavy model complexity, the contribution aligns with current consensus to deliver clinically useful, explainable, and deployable PPG diagnostics [37,38,52,56,58].

6.1. Dataset limitations and external validation

Multi-source assembly improves diversity but remains modest relative to real-world heterogeneity across demographics, comorbidities, and sensors. Domain shift from fingertip/earlobe to wrist (and from clinical to free-living capture) can alter morphology and quality, impacting generalization [30,37,45,62]. External, multi-center validation with device-diverse cohorts and standardized protocols is required, echoing recent multicenter AF screening and meta-analyses for PPG BP estimation that call for harmonized reporting and calibration assessment [35,37,38]. Establishing open, labeled, multi-endpoint datasets with class balance and subject-wise splits would materially advance comparability and clinical credibility [30,37,45].

6.2. Interpretability and the “black box” challenge

High accuracy is necessary but insufficient; clinicians require insight into which waveform regions drive decisions (e.g., systolic slope, notch timing, diastolic decay). Physics-aware notch identification and integrated explainability on scalograms or PPG matrices provide interpretable linkage between morphology and predictions and have been proved feasible with limited overhead [40,46,52]. Embedding Grad-CAM/LRP/SHAP into review workflows can expose model failure modes, support auditor review, and accelerate regulatory evaluation [46,52,58].

6.3. Integration of diverse sensing techniques

PPG is indirect and sensitive to motion, contact pressure, and illumination. Sensor fusion with accelerometry (artifact modeling), ECG (electrical rhythm confirmation), and temperature (vascular tone) is a logical extension supported by literature on wearable pipelines and HRV monitoring at the edge [15,36,51,63]. Where feasible, multi-wavelength PPG can strengthen vascular specificity and BP surrogates, as shown by hybrid learning approaches that combine spectral features with compact CNN backbones [13,37,64].

6.4. Challenges related to regulations, ethics, and implementation

Regulatory pathways require evidence of safety, effectiveness, calibration, and fairness across subgroups (age, sex, skin tone), as emphasized by consensus guidance on arrhythmia monitoring and post-stroke surveillance [58]. Prospective, stratified validation and fairness analyses are necessary to avoid performance gaps; continuous post-market monitoring with SQI logging and explainability snapshots can support surveillance and bias auditing [43,58,65]. Implementation should also consider energy budgets, latency, and privacy—areas where cloud–edge collaboration and tiny artifact detectors are increasingly practical [15,44,51].

6.5. A Strategy for progression

Priority actions include: (i) external, multi-center validation spanning demographics and hardware with standardized acquisition and subject-wise splits; (ii) native integration of XAI (Grad-CAM/LRP/SHAP) and physics-aware markers (e.g., notch) to enhance trust and discover physiology [40,46,52]; (iii) sensor-fusion prototypes (PPG + ACC + ECG±Temp) with energy-aware scheduling for wearables; and (iv) decision-curve analysis and calibration reporting to document clinical net benefit alongside accuracy [18,38,58]. Together with cloud–edge frameworks and lightweight artifact screening, these steps can convert a high-performing prototype into an auditable, equitable, and energy-efficient clinical decision tool [15,51,63].