DACIS: Disease-Aware Channel Scoring

• DACIS is a methodology that computes disease-sensitive importance scores for neural network channels using gradient norms, feature variance, and Fisher discriminant ratios.
• It integrates adaptive recalibration and meta-learning to optimize channel pruning in both clinical time series and plant pathology applications.
• Experimental benchmarks show that DACIS achieves high compression and real-time inference with minimal accuracy loss in resource- and data-constrained settings.

Disease-Aware Channel Importance Scoring (DACIS) is a principled methodology for quantifying the contribution of individual neural network channels to disease prediction and classification tasks under resource and data constraints. DACIS advances channel selection and model interpretability by directly linking each channel’s influence to the discriminative requirements of specific diseases, ensuring that retained network components preserve clinically or pathologically salient features. DACIS operates as either an adaptive recalibration mechanism (as in AICare for clinical time series (Ma et al., 2023)) or a composite importance metric for network pruning (as in the PMP pipeline for few-shot plant pathology (Alam et al., 5 Jan 2026)).

1. Mathematical Construction of DACIS

DACIS assigns to every channel in a given feature extraction architecture an explicit importance score that is disease-sensitive. In convolutional neural networks, this score at layer $\ell$, channel $c$ is formulated as:

$$\mathrm{DACIS}_{\ell}^{(c)} = \lambda_1 \mathcal{G}_{\ell}^{(c)} + \lambda_2 \mathcal{V}_{\ell}^{(c)} + \lambda_3 \mathcal{D}_{\ell}^{(c)}$$

where:

• $\mathcal{G}_{\ell}^{(c)}$: Gradient-norm, quantifying each channel's impact on loss reduction.
• $\mathcal{V}_{\ell}^{(c)}$: Feature variance after global average pooling, proxying information-carrying capacity.
• $\mathcal{D}_{\ell}^{(c)}$: Fisher discriminant ratio, measuring between- and within-disease class separability.
• $\lambda_1 + \lambda_2 + \lambda_3 = 1$, with empirical weights $(0.3, 0.2, 0.5)$ in plant pathology.

In clinical time-series models (AICare), DACIS is realized via channel-wise attention coefficients $\alpha_n$, computed on embedded feature vectors and adaptively recalibrated according to health context embeddings:

$$\begin{aligned} & f_s = \frac{1}{N+1} \sum_{i=0}^N f_i \ & q = W_s f_s,\quad k_n = W_n f_n \ & \zeta_n = q^T k_n \ & \alpha_n = \frac{e^{\zeta_n}}{\sum_{j=1}^N e^{\zeta_j}} \ & f^*_n = \alpha_n f_n \end{aligned}$$

These attention weights act as dynamic importance scores, modulated by patient disease context at each visit.

2. Pipeline Architectures Leveraging DACIS

DACIS is used as the central criterion within multi-stage compression and adaptation frameworks, the most notable of which is the Prune-then-Meta-Learn-then-Prune (PMP) pipeline for few-shot plant disease detection (Alam et al., 5 Jan 2026):

• Stage 1 (Initial Pruning): DACIS scores computed per channel. Bottom 40% pruned from base model (e.g., ResNet-18).
• Stage 2 (Meta-Learning): Episodic meta-training on few-shot tasks; meta-gradients accumulated per channel.
• Stage 3 (Refinement Pruning): Importance scores refined as $$\widetilde{\mathrm{DACIS}}_{\ell}^{(c)} = \mathrm{DACIS}_{\ell}^{(c)} (1 + \gamma |G_{\text{meta},\ell}^{(c)}|)$$, enabling further pruning to reach high compression (>78%).

AICare (Ma et al., 2023) uses a SE-style recalibration block: multi-channel Bi-GRU embeddings are attention-reweighted by $\alpha_n$ and then passed to a classification head, yielding patient- and visit-specific prediction with interpretable channel importances.

3. Disease Awareness and Context Adaptation

DACIS is characterized by explicit integration of disease structure into channel scoring:

• The Fisher discriminant $\mathcal{D}_{\ell}^{(c)}$ preserves channels critical for separating disease classes—pathogen-specific visual cues, clinically differentiated lab changes, etc.
• In AICare, dynamic weights $\alpha_n$ are conditioned on real-time health context $f_s$, directly coupling channel scaling to the patient’s disease trajectory.
• Meta-gradient refinement in PMP biases channel retention to those facilitating rapid adaptation to new disease categories under few-shot constraints.

Standard magnitude or generic sensitivity-based pruning mechanisms lack this directed, disease-conditional feature selection.

4. Interpretability and Clinical/Field Deployment

DACIS’s channel weights are directly interpretable in terms of disease relevance at prediction time:

• For AICare: Heatmaps of $\alpha_n$ across patient populations stratified by cause of death, and feature-importance trajectories over time ("turning curves"); interactive visualization tools plot risk curves and corresponding feature attentions (Ma et al., 2023).
• For PMP–DACIS: Retained channels can be traced back to activation maps, supporting downstream explainability (e.g., pixel-level Grad-CAM localization of salient disease cues) (Alam et al., 5 Jan 2026).

Ante-hoc interpretability is thus intrinsic to the DACIS methodology, not bolted on as a post hoc explanation.

5. Experimental Benchmarks and Efficiency

DACIS delivers quantifiable performance improvements, especially in resource- and data-constrained regimes:

Model	Params	Compression	5-shot Acc.	Inference FPS
Full ResNet-18	11.2 M	0%	84.6%	1.95
PMP–DACIS	2.5 M	78%	83.2%	7.0

On Raspberry Pi 4, PMP–DACIS maintains 98.4% of original accuracy, with real-time inference at 7 FPS (Alam et al., 5 Jan 2026). In medical prediction (AICare), the SE-based recalibration improves AUPRC by +11.8% and AUROC by +3.6% over baselines for peritoneal and hemodialysis mortality (Ma et al., 2023). Ablation studies consistently show substantial degradation in performance if disease-aware channel scoring is removed.

6. Limitations, Generalizability, and Future Directions

Limitations of DACIS include dependence on explicit disease taxonomies (required for Fisher scoring), lack of runtime channel-width adaptability, and elevated initial computational requirements (e.g., meta-training and task-specific pruning performed offline on GPUs). Application to novel domains may require new class hierarchies or unsupervised discriminant approximations.

Generalizable to any domain where inter-class visual or feature structure is salient (medical imaging, defect detection), DACIS naturally complements Squeeze-and-Excitation and dynamic channel gating architectures. A plausible implication is synergistic integration with quantization or neural-architecture search for joint latency–accuracy–memory optimization.

Future research avenues include continual few-shot learning, federated pruning across heterogeneous edge devices, multimodal fusion for disease diagnosis, and refined interpretability via mapping DACIS-retained channels to input regions and causal patterns.

7. Significance in Resource-Constrained and Data-Limited Environments

DACIS operationalizes disease specificity and discriminative power in feature ranking for channel pruning and importance recalibration. By efficiently encoding and preserving key disease features, it enables high-accuracy, interpretable, and real-time prediction and classification on devices such as Raspberry Pi 4, supporting practical deployment in field settings and clinical environments where both data and computational budgets are constrained (Alam et al., 5 Jan 2026). In longitudinal clinical modeling, individualized disease-aware importance scoring facilitates personalized medicine and early intervention guided by transparent feature trajectories (Ma et al., 2023).