Explainable Dual-Omics Filtering (EDOF)

• Explainable Dual-Omics Filtering (EDOF) is a framework that integrates spatial dosiomic and radiomic features to achieve transparent, voxel-level radiation pneumonitis prediction in breast cancer patients.
• The methodology employs a two-pronged filtering approach that extracts 30 dosiomic features capturing local dose intensity and 5 radiomic features detailing CT-based tissue heterogeneity, which are then integrated using an Explainable Boosting Machine (EBM).
• The framework reported strong performance with an AUC of 0.95 ± 0.01, sensitivity of 0.81 ± 0.05, and specificity of 0.93 ± 0.01, while providing interpretable partial dependence plots to map risk thresholds.

Searching arXiv for the specified paper to ground the article in the source text. arxiv_search(query="(Yang et al., 4 Aug 2025)", max_results=5) arXiv search results for "(Yang et al., 4 Aug 2025)":

1. id: (Yang et al., 4 Aug 2025) title: A Dual Radiomic and Dosiomic Filtering Technique for Locoregional Radiation Pneumonitis Prediction in Breast Cancer Patients authors: ['YI-CHEN SHEN', 'KO-WEN WU', 'FU-CHUN HUANG', 'JING-HAO HUANG', 'SAN-YUAN HUANG', 'HUI-TING CHEN', 'YAO-LUNG KANG'] abstract: Purpose: Radiation pneumonitis (RP) is a serious complication of intensity-modulated radiation therapy (IMRT) for breast cancer patients, underscoring the need for precise and explainable predictive models. This study presents an Explainable Dual-Omics Filtering (EDOF) model that integrates spatially localized dosiomic and radiomic features for voxel-level RP prediction. Methods: A retrospective cohort of 72 breast cancer patients treated with IMRT was analyzed, including 28 who developed RP. The EDOF model consists of two components: (1) dosiomic filtering, which extracts local dose intensity and spatial distribution features from planning dose maps, and (2) radiomic filtering, which captures texture-based features from pre-treatment CT scans. These features are jointly analyzed using the Explainable Boosting Machine (EBM), a transparent machine learning model that enables feature-specific risk evaluation. Model performance was assessed using five-fold cross-validation, reporting area under the curve (AUC), sensitivity, and specificity. Feature importance was quantified by mean absolute scores, and Partial Dependence Plots (PDPs) were used to visualize nonlinear relationships between RP risk and dual-omic features. Results: The EDOF model achieved strong predictive performance (AUC = 0.95 ± 0.01; sensitivity = 0.81 ± 0.05). The most influential features included dosiomic Intensity Mean, dosiomic Intensity Mean Absolute Deviation, and radiomic SRLGLE. PDPs revealed that RP risk increases beyond 5 Gy and rises sharply between 10–30 Gy, consistent with clinical dose thresholds. SRLGLE also captured structural heterogeneity linked to RP in specific lung regions. Conclusion: The EDOF framework enables spatially resolved, explainable RP prediction and may support personalized radiation planning to mitigate pulmonary toxicity. Need write article. Since user insists every concrete claim must appear verbatim in data block. Need be careful. Could include only facts present. We can synthesize structure but factual claims should be from provided data. Since details include more exact claims beyond abstract. We can mention title and authors maybe from search. User says reference authors/groups when mentioned in data. Authors weren't in provided data but in arxiv search result. Since general guidelines require use search tool, we can cite authors maybe. But to be safe maybe mention "reported by Yi-Chen Shen, Ko-Wen Wu, Fu-Chun Huang, Jing-Hao Huang, San-Yuan Huang, Hui-Ting Chen, and Yao-Lung Kang" from search. Since concrete claim from search. Fine.

Need 5-7 numbered h2 sections. Numbered in markdown like ## 1. Definition and scope etc.

Need no heading at beginning. Start with intro paragraph. Neutral tone.

Need perhaps 6 sections: 1 clinical problem and aim, 2 architecture/components, 3 dosiomic filtering, 4 radiomic filtering, 5 EBM and explainability, 6 performance and clinical significance. Could also include implementation and limitations/interpretive notes. Need comprehensive coverage.

Need maybe mention misconceptions/controversies objectively. Since only data mentions black-box models, deep learning low sensitivity, scenarios for model failure, out-of-field RP. Could say "A common misconception would be to treat EDOF as a purely dosimetric surrogate; the framework jointly models dose-plan features and CT-derived texture features." Since this is an interpretation, need mark "This suggests" etc. Need keep factual info first, then contextual significance.

Need avoid invented tool names not in data. Need if mention "dual-omics" as Editor's term? no need.

Need include formulas exactly as given. Fine.

Need perhaps a small table summarizing components maybe okay. But not necessary. Keep paragraphs. Could include a table with component/input/output, all from data. But they said tables only organize info already present. It's okay. But maybe avoid to minimize risk.

Need cite (Yang et al., 4 Aug 2025) in intro and perhaps once per section at most. Since primarily single paper, cite introduction and specific claims per section. The rule says cite given paper at most once per section. okay.

Need be careful that specificity 0.93 ± 0.01 and DSC were in details not abstract, so okay to include as provided. Need note "Comparisons: only radiomics AUC 0.68; only dosiomics AUC 0.94; GBT AUC=0.95, sensitivity slightly lower; U-Net++ sensitivity 0.00." These are in data. Great.

Need mention "custom MATLAB toolbox" and "InterpretML Python library" from details. okay.

Need ensure no mechanical repetition. Maybe cite intro and section 5 and 6.

Let's craft concise but comprehensive. Explainable Dual-Omics Filtering (EDOF) is a machine learning framework for predicting locoregional risk of radiation pneumonitis (RP) after breast cancer radiotherapy by generating explainable, voxel-level risk maps rather than only patient-level risk estimates. Reported in “A Dual Radiomic and Dosiomic Filtering Technique for Locoregional Radiation Pneumonitis Prediction in Breast Cancer Patients” by Yi-Chen Shen, Ko-Wen Wu, Fu-Chun Huang, Jing-Hao Huang, San-Yuan Huang, Hui-Ting Chen, and Yao-Lung Kang, the framework combines dosiomic filtering from 3D dose distributions, radiomic filtering from pre-treatment CT, and the Explainable Boosting Machine (EBM) for interpretable prediction (Yang et al., 4 Aug 2025). In the reported study, EDOF was developed for breast cancer patients treated with intensity-modulated radiation therapy (IMRT), with the stated aim of precise and explainable prediction of RP, a serious complication of treatment.

1. Clinical problem and modeling objective

Radiation pneumonitis is presented as a serious complication of IMRT for breast cancer patients, and the framework is motivated by the need for precise and explainable predictive models. The study analyzed a retrospective cohort of 72 breast cancer patients treated with IMRT, including 28 who developed RP (Yang et al., 4 Aug 2025).

The framework is explicitly designed for spatially resolved prediction. Rather than restricting inference to global patient-level endpoints, it predicts risk at the voxel level across the lung. This spatial formulation is central to the model’s intended clinical role: it identifies not just overall dose burden, but which spatial regions and tissue types, as characterized by texture, are the most vulnerable. A plausible implication is that EDOF occupies an intermediate position between conventional dose-volume summarization and opaque end-to-end image prediction, because it retains localized feature construction while preserving feature-level interpretability.

2. Core architecture of the EDOF framework

EDOF consists of three stated components: dosiomic filtering, radiomic filtering, and integration and prediction with EBM (Yang et al., 4 Aug 2025). The first component extracts spatially localized dose features from treatment dose distributions. The second captures texture-based features remotely and locally from pre-treatment CT images. The third combines both feature sets in a transparent, interpretable model.

The architecture is organized around local neighborhood analysis. For both dose maps and CT scans, a 3D isotropic kernel of size $7\times7\times7\ \text{mm}^3$ is swept across each voxel of the lung, forming local patches. Each voxel is therefore represented by a feature vector composed of filtered dosiomic and radiomic descriptors. After feature reduction, the reported representation contains 30 dosiomic features and 5 radiomic features.

This design emphasizes dual-omic integration in a strictly localized setting. The paper characterizes dosiomic filtering as capturing local dose intensity and spatial distribution, while radiomic filtering captures local tissue heterogeneity, density, and textural characteristics. This suggests that EDOF is intended to model RP risk as a function of both delivered radiation pattern and pre-treatment tissue state, rather than treating dose alone as sufficient.

3. Dosiomic filtering

The stated purpose of dosiomic filtering is to capture local spatial statistics and intensity characteristics of the planned radiation dose beyond standard dose-volume indices (DVIs). Its inputs are 3D dose distribution maps and segmented lung volumes. For each local patch, a set of 19 candidate dosiomic features is computed, describing dose intensity, variation, skewness, and higher-order statistics.

Feature reduction is performed to avoid collinearity and redundancy. The reported procedure computes Pearson correlation for all feature pairs, applies hierarchical clustering to the merged correlation matrix, and selects one representative within each feature cluster, resulting in a final set of 30 independent dosiomic features (Yang et al., 4 Aug 2025).

The dosiomic filtering operation is expressed as

$$D^n_{i,j,k} = \text{Feature}_n(\text{dose values in neighborhood of } (i,j,k))$$

for each voxel $(i,j,k)$ in an $I\times J\times K$ lung grid and for each dosiomic feature $n$, where the neighborhood is defined by the $7\times7\times7\ \text{mm}^3$ window.

Among the features identified as most influential are dosiomic Intensity Mean and dosiomic Intensity Mean Absolute Deviation. The former is described as local mean dose in Gy and is stated to capture absolute exposure, directly mapping to key clinic metrics like V5Gy and V20Gy. The latter captures local dose variability, including edges of dose gradients or heterogeneous exposure zones. In the framework’s interpretation, these quantities contribute substantially to prediction accuracy through their normalized mean absolute scores.

4. Radiomic filtering

Radiomic filtering is intended to encode local tissue heterogeneity, density, and textural characteristics from CT images, motivated by the claim that these are implicated in tissue susceptibility to RP. The inputs are pre-treatment CT images and the lung mask. As in the dosiomic branch, the same $7\times7\times7\ \text{mm}^3$ sliding window is used.

For each patch, 53 radiomic features are evaluated. The reported families include GLCM, which measures texture such as homogeneity and contrast; GLRLM, which captures runs of similar intensity and thereby granularity; and GLSZM, which assesses size and uniformity of homogeneous zones. The same collinearity reduction strategy is then applied, yielding 5 independent radiomic features (Yang et al., 4 Aug 2025).

The radiomic filtering operation is written as

$$R^m_{i,j,k} = \text{Feature}_m(\text{CT intensities in neighborhood of } (i,j,k))$$

for each radiomic feature $m$. After joint construction with the dosiomic branch, each voxel is represented by a vector of $p+q$ features, specified in the report as 30 dosiomic plus 5 radiomic features.

A prominent radiomic variable in the reported model is GLRLM Short Run Low Gray Level Emphasis (SRLGLE). The paper states that SRLGLE reflects local structural heterogeneity and that high values mean more small, low-density regions, possibly indicating vulnerability to RP. The abstract further states that SRLGLE captured structural heterogeneity linked to RP in specific lung regions (Yang et al., 4 Aug 2025). A plausible implication is that radiomic filtering supplies susceptibility information not directly recoverable from dose distributions alone.

5. Explainable prediction with EBM

The predictive model is the Explainable Boosting Machine, described as leveraging generalized additive models (GAMs) and fitting independent, flexible, tree-based response functions to each feature. For a voxel feature vector $$D^n_{i,j,k} = \text{Feature}_n(\text{dose values in neighborhood of } (i,j,k))$$0, the prediction model is given by

$$D^n_{i,j,k} = \text{Feature}_n(\text{dose values in neighborhood of } (i,j,k))$$1

where $$D^n_{i,j,k} = \text{Feature}_n(\text{dose values in neighborhood of } (i,j,k))$$2 is the logit of the predicted probability for binary classification, $$D^n_{i,j,k} = \text{Feature}_n(\text{dose values in neighborhood of } (i,j,k))$$3 is a learned function for feature $$D^n_{i,j,k} = \text{Feature}_n(\text{dose values in neighborhood of } (i,j,k))$$4 via boosting, and $$D^n_{i,j,k} = \text{Feature}_n(\text{dose values in neighborhood of } (i,j,k))$$5 is the intercept (Yang et al., 4 Aug 2025).

The explainability mechanisms reported in the study are feature importance quantified by mean absolute contribution of each $$D^n_{i,j,k} = \text{Feature}_n(\text{dose values in neighborhood of } (i,j,k))$$6, Partial Dependence Plots (PDPs), and voxel-wise class predictions. PDPs visualize nonlinear relationships between RP risk and dual-omic features, while voxel-wise class predictions permit spatial risk maps over the lung. The paper characterizes EBM as transparent and interpretable, with explicit visualization of how risk is built from each feature.

Three top predictive features are highlighted after filtering and clustering: dosiomic Intensity Mean, radiomic GLRLM SRLGLE, and dosiomic Intensity Mean Absolute Deviation. The corresponding PDP findings are described in detail. For Intensity Mean, RP risk is minimal at doses below 5 Gy, rises sharply between 8 and 30 Gy, and plateaus at very high doses; the abstract summarizes this as risk increasing beyond 5 Gy and rising sharply between 10–30 Gy, consistent with clinical dose thresholds (Yang et al., 4 Aug 2025). For SRLGLE, the relationship is complex, with high risk at both very low and moderate values, suggestive of texture-density interactions. For Intensity Mean Absolute Deviation, risk rises sharply up to approximately 1 Gy and then saturates.

These reported PDPs are central to the framework’s claim of clinical interpretability. The paper states that they allow clinicians to see risk thresholds and dose-response relationships, supporting clinical defensibility and trust. It also states that examining PDPs and importance can identify cases where predictions are less reliable, for example out-of-field RP. This suggests that explainability is treated not only as a communication layer but also as a mechanism for model auditing.

6. Reported performance, comparisons, and implementation

Model performance was assessed using five-fold cross-validation, reporting area under the curve (AUC), sensitivity, and specificity. The reported EDOF performance is AUC $$D^n_{i,j,k} = \text{Feature}_n(\text{dose values in neighborhood of } (i,j,k))$$7, sensitivity $$D^n_{i,j,k} = \text{Feature}_n(\text{dose values in neighborhood of } (i,j,k))$$8, and specificity $$D^n_{i,j,k} = \text{Feature}_n(\text{dose values in neighborhood of } (i,j,k))$$9; for Grade II RP patients, the reported DSC is $(i,j,k)$0 (Yang et al., 4 Aug 2025).

The paper also reports several comparisons. Using only radiomics without dose yields AUC $(i,j,k)$1. Using only dosiomics yields AUC $(i,j,k)$2, described as almost as good but with lower sensitivity than dual-omics. Dual-omics with classic GBT yields AUC $(i,j,k)$3, with sensitivity stated to be slightly lower and interpretability much lower than EBM. Deep learning with U-Net++ is reported to show low sensitivity $(i,j,k)$4, which the source attributes to challenges with imbalanced or small RP regions.

These comparisons delimit the framework’s claimed contribution. The results are presented as evidence that the radiomic branch alone is insufficient, the dosiomic branch is strong but incomplete, and the dual-omics EBM formulation preserves discrimination while improving interpretability relative to black-box alternatives. A plausible implication is that the marginal value of radiomics in this setting lies less in replacing dose information than in refining susceptibility estimates within dose-exposed regions.

The technical implementation is also specified. Voxel-level features were generated via a custom MATLAB toolbox for radiomic and dosiomic filtering, and EBM was fitted using the Microsoft InterpretML Python library, with careful cross-validation and early stopping to prevent overfitting (Yang et al., 4 Aug 2025).

7. Clinical significance and scope of interpretation

The paper concludes that the EDOF framework enables spatially resolved, explainable RP prediction and may support personalized radiation planning to mitigate pulmonary toxicity (Yang et al., 4 Aug 2025). Its stated clinical relevance includes direct mapping of PDPs for Intensity Mean to clinic standards such as V5, V10, and V20 dose buckets; identification of spatial regions and tissue types most vulnerable to RP; and generation of voxel-level risk maps that may be useful for treatment adaptation or local sparing.

The framework is also presented as improving explainability and utility through direct mapping to clinical practice, transparent model behavior, clinician-valid explanations, and spatial resolution. The source contrasts this with prior patient-level black-box models and emphasizes that EDOF can highlight avoidance areas for personalized radiation planning. Because the central explanatory variables include Intensity Mean, Intensity Mean Absolute Deviation, and SRLGLE, the framework links risk to both local dose burden and local structural heterogeneity.

Several interpretive cautions follow from the reported design. The study is based on a retrospective cohort of 72 patients, and the performance claims are tied to five-fold cross-validation within that dataset. The paper also notes scenarios where predictions are less reliable, such as out-of-field RP. This suggests that EDOF should be understood as an explainable locoregional prediction framework whose strongest claims concern within-cohort voxel-level RP mapping and clinically interpretable dose-texture associations, rather than unrestricted generalization beyond the reported setting.