PEFRF: Pulmonary & Image Quality Frameworks

• PEFRF is an acronym representing two distinct frameworks: one for non-contact pulmonary function prediction and one for full-reference image quality assessment.
• In pulmonary monitoring, the framework integrates multimodal video data and patient metadata using lightweight CNNs and SNNs to predict PEF and FEV1/FVC ratios with high precision.
• In image quality assessment, PEFRF fuses permutation entropy features with a Random Forest regressor, achieving top performance across diverse natural and screen-content image datasets.

Searching arXiv for the cited PEFRF-related papers to ground the article in current literature. PEFRF is an acronym that denotes two unrelated 2025 research frameworks in arXiv literature. In pulmonary health monitoring, it refers to a Peak Expiratory Flow Regression Framework embedded in PulmoFusion, a multimodal, non-contact spirometry system that predicts PEF and the FEV1/FVC ratio from RGB or thermal video plus patient metadata (Sharshar et al., 29 Jan 2025). In image quality assessment, it denotes Permutation Entropy-based Features Fused with Random Forest, a full-reference IQA method that combines permutation-entropy features from gradient maps with a Random Forest regressor to assess both natural-scene images and screen-content images (Baqar et al., 24 Aug 2025). The shared acronym is therefore polysemous rather than indicative of a common technical lineage.

1. Nomenclature and scope

The pulmonary usage appears in "PulmoFusion: Advancing Pulmonary Health with Efficient Multi-Modal Fusion" (Sharshar et al., 29 Jan 2025). There, PEFRF is defined as the overall regression framework that predicts Peak Expiratory Flow (PEF), and also predicts FEV1/FVC from multimodal non-contact data. The system further includes an FEV1/FVC abnormality detector, so the acronym denotes only the regression component in a broader remote spirometry pipeline.

The image-quality usage appears in "A Hybrid Approach for Unified Image Quality Assessment: Permutation Entropy-Based Features Fused with Random Forest for Natural-Scene and Screen-Content Images for Cross-Content Applications" (Baqar et al., 24 Aug 2025). There, PEFRF is defined explicitly as Permutation Entropy-based Features Fused with Random Forest, a full-reference image quality assessment (FR-IQA) framework intended as a single unified metric across natural-scene images (NSIs) and screen-content images (SCIs).

A plausible implication is that any technical discussion of "PEFRF" requires immediate domain disambiguation, since the acronym refers to fundamentally different inputs, objectives, and evaluation criteria in the two works.

2. PEFRF in remote pulmonary monitoring

In PulmoFusion, PEFRF addresses the limitations of traditional remote spirometry identified in the paper: cost and accessibility barriers, hygiene and infection risk from mouthpieces and airflow sensors, limited generalization and personalization in prior remote approaches, and the high computational cost of many deep-learning solutions for video or thermal data (Sharshar et al., 29 Jan 2025). The framework targets non-contact RGB/thermal video of the face/mask, rich patient metadata, energy-efficient Spiking Neural Networks (SNNs) for classification, lightweight video CNNs (X3D) for regression, Multi-Head Attention (MHA) for multimodal fusion, and K-fold cross-validation and ensemble learning for robustness.

The pulmonary function measures are explicitly separated into regression and classification tasks. For PEF, the task is regression from video plus metadata, with ground truth from a Peak Flow Asthma Meter. For FEV1 and FVC, the values are estimated from NHANES III reference equations, and their ratio is used both as a continuous regression target and as a binary clinical classification target with threshold

$$\text{FEV1/FVC (measured/predicted)} \geq 70\%\ \Rightarrow \text{normal},\quad <70\% \Rightarrow \text{abnormal}.$$

The data source is the RespiroDynamics dataset, comprising 60 volunteers aged 15–75 years, with two sessions per subject, one resting and one post-exercise, each of approximately 1 minute, and with both RGB and thermal recordings (Sharshar et al., 29 Jan 2025). The collected signals include RGB video, thermal video, heart rate, smartwatch ECG, blood pressure, PEF, and detailed metadata such as smoking duration, athlete status, seasonal cough, lung past/genetic problems, and InBody data. Videos were manually inspected and cropped to isolate the face region, then segmented into 2,424 cycle-level clips corresponding to individual breathing cycles.

The practical significance of this PEFRF is its non-contact and hygienic design, its compatibility with at-home monitoring for asthma or COPD, and its explicit modeling of patient-specific context through metadata. The paper further states that SNN-based classification is substantially lighter than CNN training, citing 4 GB GPU for SNN classification versus 24 GB RAM (Quadro RTX 6000) for the CNN regression model, and inference speeds of approximately 0.2 s per patient for SNN and 1.3 s per patient for CNN (Sharshar et al., 29 Jan 2025).

3. Architecture and learning pipeline of the pulmonary PEFRF

PulmoFusion defines an end-to-end architecture from synchronized multimodal capture to pulmonary predictions (Sharshar et al., 29 Jan 2025). RGB and thermal recordings are synchronized by timestamp, the face or mask region is isolated using OpenCV after manual inspection, and each segmented breathing cycle is standardized to 30 frames at $224 \times 224$ resolution. Data augmentation is applied in spatial and temporal form. Because metadata are constant across cycles while respiratory performance varies from cycle to cycle, cycle-level predictions are averaged in a post-processing step to produce patient-wise estimates for regression.

For classification of FEV1/FVC abnormality, the framework uses SNNs with rate-coded spikes. Each video frame is converted into spike trains according to

$u_{ij} \sim \text{Bernoulli}(p_{ij}),$

where $p_{ij}$ is proportional to pixel intensity, and metadata are similarly encoded into spike trains. The central neuron model is the Leaky Integrate-and-Fire (LIF) neuron: $$m^{(\ell)}(t) = \beta\, v^{(\ell)}(t - 1) + W^{(\ell)} s^{(\ell-1)}(t),$$

$$s^{(\ell)}(t) = H\Big(m^{(\ell)}(t) - V_{\text{th}^{(\ell)}}\Big),$$

$$v^{(\ell)}(t) = m^{(\ell)}(t) - s^{(\ell)}(t)\, V_{\text{th}^{(\ell)}}.$$

Because the Heaviside step function is non-differentiable, training relies on surrogate gradients and backprop through time, with ADAM as optimizer. The CSNN architecture consists of spike encoding, 2D spiking convolutional layers, spiking pooling, flattening and spiking fully connected layers, metadata integration by concatenation, and a classification head that uses spike frequency or accumulated potentials. The RGB SNN is implemented as an ensemble of four SNN models with different convolutional filters and parameters.

For regression of PEF and FEV1/FVC, the paper adopts a hybrid design, explicitly stating that SNNs remain limited for regression tasks in critical healthcare applications (Sharshar et al., 29 Jan 2025). The regression backbone is a lightweight X3D 3D CNN processing 30-frame video clips, while metadata are passed through a fully connected layer to obtain a compact representation. Fusion is performed with Multi-Head Attention rather than simple concatenation. The formulation is written in terms of video features $F_{\text{vid}}$ and metadata features $F_{\text{meta}}$, with per-head projections

$$Q_h = F_{\text{vid}} W_h^Q,\quad K_h = F_{\text{meta}} W_h^K,\quad V_h = F_{\text{meta}} W_h^V,$$

followed by scaled dot-product attention and multi-head composition. The resulting fused representation is passed to a final fully connected regression head that outputs continuous PEF and FEV1/FVC ratio.

The learning objective for regression is Mean Squared Error (MSE), and evaluation uses RMSE, Relative RMSE (RelRMSE), and MAE. Robustness is increased by training multiple models and combining them by ensemble averaging for regression or majority voting for classification.

4. Empirical performance of the pulmonary PEFRF

The evaluation protocol in PulmoFusion uses an 80%/20% train/test split at the subject level and 5-fold cross-validation with distinct subjects in train and test partitions (Sharshar et al., 29 Jan 2025). Performance is reported both on a breathing-cycle basis and on a patient-wise basis, the latter obtained by aggregating predictions across cycles. The paper explicitly emphasizes that patient-level aggregation is clinically more relevant and substantially more accurate.

For classification, the key results show strong performance for both SNN and CNN variants. The SNN multimodal thermal model achieves 91.99% ± 2.0% cycle-based accuracy and 98.50% ± 0.5% patient-wise accuracy, with corresponding F1 scores of 91.61% ± 3.0% and 98.80% ± 0.5% (Sharshar et al., 29 Jan 2025). The CNN multimodal thermal model reaches 94.89% ± 2.0% cycle-based accuracy and 99.50% ± 0.5% patient-wise accuracy. The RGB variants are weaker at cycle level but remain very strong after patient-wise aggregation.

For PEF regression, the ablation sequence shows progressive gains from augmentation, metadata fusion, MHA, and ensemble learning. On thermal data, the single baseline gives RelRMSE = 0.30, RelMAE = 0.26, Pearson = 0.72; the final PEFRF configuration with multimodal input, MHA, and ensemble learning reaches RelRMSE = 0.11 ± 0.05, RelMAE = 0.09 ± 0.06, and Pearson = 0.93 ± 0.04 (Sharshar et al., 29 Jan 2025). On RGB data, the best configuration reaches RelRMSE = 0.26 ± 0.07, RelMAE = 0.21 ± 0.04, Pearson = 0.79 ± 0.04.

For FEV1/FVC regression, the best thermal configuration achieves RelRMSE = 0.06 ± 0.03, RelMAE = 0.05 ± 0.05, Pearson = 0.96 ± 0.01, while the best RGB configuration achieves RelRMSE = 0.12 ± 0.03, RelMAE = 0.10 ± 0.04, Pearson = 0.85 ± 0.05 (Sharshar et al., 29 Jan 2025). The abstract reports an MAE of 4.52% for FEV1/FVC predictions and states that the results establish state-of-the-art performance relative to prior remote spirometry methods cited in the paper.

These results support two central conclusions within the paper’s experimental frame: thermal video is stronger than RGB for regression, and metadata-aware attention-based fusion materially improves prediction quality over video-only or simple-concatenation baselines.

5. PEFRF in full-reference image quality assessment

In the IQA literature, PEFRF is a full-reference image quality assessment framework for cross-content applications, intended to operate reliably on both NSIs and SCIs (Baqar et al., 24 Aug 2025). The motivation is that standard FR-IQA measures such as PSNR, SSIM, MS-SSIM, and GMSD were largely developed for natural images, whereas screen-content images exhibit sharp edges, high-contrast text, large flat regions, and mixed graphics-text-photo composition. The paper characterizes the resulting failure mode as a cross-content generalization gap.

PEFRF addresses this by working in the gradient domain, extracting permutation entropy from local patches of the gradient maps of the reference, distorted, and fused images, and learning the mapping from these features to subjective scores via a Random Forest regressor (Baqar et al., 24 Aug 2025). The method is hybrid rather than purely hand-crafted or purely deep-learning-based. Its stated goals are unified performance across NSIs and SCIs, robustness to many distortion types, and good interpretability and computational efficiency relative to heavy deep models.

The framework uses grayscale normalized inputs

$224 \times 224$0

and computes Sobel-based gradient magnitudes

$224 \times 224$1

From the reference and distorted gradients, it constructs a fused gradient image

$224 \times 224$2

These three maps are denoted $224 \times 224$3, $224 \times 224$4, and $224 \times 224$5.

The central signal descriptor is normalized permutation entropy

$224 \times 224$6

with permutation order $224 \times 224$7 and delay $224 \times 224$8 (Baqar et al., 24 Aug 2025). For each of the three gradient maps, the method takes a $224 \times 224$9 patch, flattens it to a length-9 sequence, and uses a sliding 3-sample window to enumerate ordinal patterns. Counting pattern frequencies yields a patch-wise entropy value, and repeating this with stride 1 produces three permutation entropy maps: $u_{ij} \sim \text{Bernoulli}(p_{ij}),$0, $u_{ij} \sim \text{Bernoulli}(p_{ij}),$1, and $u_{ij} \sim \text{Bernoulli}(p_{ij}),$2.

6. IQA formulation, datasets, and reported results

The three entropy maps are combined into a Local Quality Map (LQM): $u_{ij} \sim \text{Bernoulli}(p_{ij}),$3 where $u_{ij} \sim \text{Bernoulli}(p_{ij}),$4 ensures numerical stability (Baqar et al., 24 Aug 2025). This map is then summarized into a three-dimensional feature vector: $u_{ij} \sim \text{Bernoulli}(p_{ij}),$5 where $u_{ij} \sim \text{Bernoulli}(p_{ij}),$6 is the mean absolute deviation around the LQM mean, $u_{ij} \sim \text{Bernoulli}(p_{ij}),$7 is the standard deviation, and $u_{ij} \sim \text{Bernoulli}(p_{ij}),$8 is the mean intensity of the LQM. These features represent overall similarity and spatial dispersion of local quality.

The final quality score is produced by a Random Forest Regressor with 200 decision trees and maximum tree depth 20, with hyperparameters tuned by grid search (Baqar et al., 24 Aug 2025). The paper writes the regression as

$u_{ij} \sim \text{Bernoulli}(p_{ij}),$9

For alignment with MOS or DMOS, a five-parameter logistic regression is then applied: $p_{ij}$0

Evaluation is performed on 13 FR-IQA benchmark datasets with over 21,000 distorted images and 300+ reference images, including NSI datasets such as LIVE, CSIQ, TID2013, KADID-10k, the SCI dataset SIQAD, and mixed-content data such as CCT (Baqar et al., 24 Aug 2025). Distortions include Gaussian noise, Gaussian blur, JPEG, JPEG2000, contrast changes, color saturation changes, and multiple-distortion settings. Metrics include SRCC, KRCC, PLCC, and RMSE, with formulas provided in the paper. Validation uses five-fold cross-validation and also leverages Out-of-Bag (OOB) error for Random Forest tuning.

The reported results place PEFRF at or near the top across the datasets and against more than 40 IQA metrics (Baqar et al., 24 Aug 2025). The paper states that across all 13 datasets, PEFRF consistently has the top SRCC and PLCC and the lowest RMSE, typically ranked 1st. Reported examples include LIVE: SRCC = 0.9452, PLCC = 0.9523; CSIQ: SRCC = 0.9636, PLCC = 0.9651; TID2013: SRCC = 0.9603, PLCC = 0.9681; and SIQAD: SRCC = 0.9070, PLCC = 0.9100 (Baqar et al., 24 Aug 2025). On KADID-10k, the paper reports PLCC ≈ 0.9813 and SRCC ≈ 0.9799, outperforming deep methods such as DISTS and PieAPP. The authors further report F-test evidence, for example on CIDIQ, indicating statistically significant superiority over most competing metrics.

7. Comparative interpretation, limitations, and future directions

The two PEFRF frameworks share an acronym but not a methodology. The pulmonary PEFRF is a multimodal physiological regression framework built around X3D, MHA, metadata integration, and SNN-assisted abnormality classification (Sharshar et al., 29 Jan 2025). The IQA PEFRF is a full-reference perceptual quality metric built around Sobel gradients, local permutation entropy, a tri-map LQM, and Random Forest regression (Baqar et al., 24 Aug 2025). One predicts pulmonary measurements from temporal video and patient covariates; the other predicts subjective image quality from paired reference and distorted images.

Their commonality is mainly architectural at a high level: both fuse heterogeneous signals and then learn a regression mapping. This suggests a generic design pattern—compact structural descriptors plus supervised regression or multimodal feature fusion—but the papers do not claim any direct relation between the two uses of PEFRF.

Both works also state clear limitations. In PulmoFusion, the main constraints are dataset size and diversity with only 60 subjects, reliance on high-quality, manually segmented datasets, the need for automated preprocessing techniques, sensitivity of RGB to lighting and occlusion, and the immaturity of SNN regression for critical healthcare tasks (Sharshar et al., 29 Jan 2025). Future directions include larger and more diverse cohorts, automated face/mask detection, automated breathing-cycle segmentation, improved SNN regression, model compression, neuromorphic deployment, and integration with other physiological signals.

In the IQA PEFRF, limitations include full-reference dependence, the computational overhead of permutation entropy on sliding windows relative to simple metrics, possible saturation under extreme distortions, fixed PE parameters $p_{ij}$1, and the current single-scale, single-channel grayscale formulation (Baqar et al., 24 Aug 2025). Proposed extensions include no-reference or reduced-reference variants, integration with deep learning, multi-scale or multi-channel PE, video quality assessment, and domain-specific adaptations such as medical imaging or satellite imagery.

Taken together, the acronym PEFRF in 2025 arXiv usage denotes two domain-specific frameworks that are technically unrelated yet each seek a unified solution within its own problem class: remote non-contact pulmonary function estimation in one case, and cross-content full-reference image quality assessment in the other (Sharshar et al., 29 Jan 2025, Baqar et al., 24 Aug 2025).