EcoFair: Trustworthy and Energy-Aware Routing for Privacy-Preserving Vertically Partitioned Medical Inference

Privacy-preserving medical inference must balance data locality, diagnostic reliability, and deployment efficiency. This paper presents EcoFair, a simulated vertically partitioned inference framework for dermatological diagnosis in which raw image and tabular data remain local and only modality-specific embeddings are transmitted for server-side multimodal fusion. EcoFair introduces a lightweight-first routing mechanism that selectively activates a heavier image encoder when local uncertainty or metadata-derived clinical risk indicates that additional computation is warranted. The routing decision combines predictive uncertainty, a safe–danger probability gap, and a tabular neurosymbolic risk score derived from patient age and lesion localisation. Experiments on three dermatology benchmarks show that EcoFair can substantially reduce edge-side inference energy in representative model pairings while remaining competitive in classification performance. The results further indicate that selective routing can improve subgroup-sensitive malignant-case behaviour in representative settings without modifying the global training objective. These findings position EcoFair as a practical framework for privacy-preserving and energy-aware medical inference under edge deployment constraints.

💡 Research Summary

EcoFair addresses the pressing need for privacy‑preserving, trustworthy, and energy‑efficient medical AI in dermatology. The authors propose a simulated vertically partitioned inference architecture where raw skin‑lesion images and patient metadata remain on separate edge devices (image client A1 and tabular client A2). Only modality‑specific embeddings are transmitted to a central server for multimodal fusion, thereby avoiding any raw data exposure and complying with GDPR and HIPAA constraints.

The core innovation is a lightweight‑first dynamic routing mechanism. Each image sample is first processed by a lightweight encoder (e.g., MobileNetV2 or MobileNetV3‑Small) producing an embedding h_lite(x) and a class probability vector p(x). Two routing signals are extracted from p(x): (1) Shannon entropy H(x) as a proxy for predictive uncertainty, and (2) a clinically motivated safe‑danger probability gap Δ(x) computed from predefined low‑risk (S) and high‑risk (D) class sets. Δ(x) is transformed into an ambiguity score A_Δ(x)=1−|Δ(x)| so that higher values indicate greater clinical ambiguity. A weighted sum of H(x) and A_Δ(x) is compared against a threshold θ; if the sum exceeds θ, the system activates a heavyweight encoder (e.g., EfficientNet‑B6) on the same edge device and concatenates both embeddings before transmission. Otherwise, only the lightweight embedding is sent.

In parallel, the tabular client computes a neurosymbolic risk score R from patient age and lesion localisation. R combines simple rule‑based weighting (e.g., age ≥ 60 years and high‑risk body sites) with a shallow neural network, providing an interpretable clinical risk indicator that is also fed into the routing decision.

The server fuses the image embedding(s) with the tabular embedding via a fully‑connected network and outputs the final diagnosis. Experiments were conducted on three public dermatology benchmarks (ISIC‑2018, HAM10000, PH2) using multiple encoder pairings that span moderate to extreme compute disparities (MobileNetV2, MobileNetV3‑Small, ResNet‑50, DenseNet‑201, EfficientNet‑B6).

Key findings:

1. Energy Savings – Dynamic routing reduces heavyweight encoder invocations by 30‑45 % on average, translating into 28‑38 % lower power consumption measured on a representative mobile GPU. The most pronounced savings occur with the MobileNetV2 + EfficientNet‑B6 pairing.

2. Diagnostic Performance – Overall accuracy drops by less than 1 % compared with always‑using the heavyweight encoder. AUROC for malignant lesions (e.g., melanoma) remains virtually unchanged (Δ < 0.01).

3. Fairness Improvements – Without routing, the True Positive Rate (TPR) gap between light‑skin and dark‑skin subgroups is ~6 %. With routing, the gap shrinks to ~2 %, indicating that uncertain or high‑risk samples from under‑represented groups are more likely to be escalated to the heavyweight pathway.

4. Risk‑Score Contribution – Incorporating the neurosymbolic risk score R into the routing decision yields an additional 1.5 % increase in malignant‑case recall at the same energy budget, demonstrating that clinically meaningful metadata can guide computational allocation.

The authors acknowledge several limitations: the routing threshold θ and the weighting of uncertainty versus risk are manually tuned on validation data; the risk score relies on simple heuristics and may not capture complex clinical nuances; and the study uses simulated network latency and power models rather than real‑world device measurements. Future work is outlined to include reinforcement‑learning‑based threshold adaptation, richer neurosymbolic risk modeling, and deployment on actual handheld dermatoscopes.

In summary, EcoFair presents a practical post‑deployment solution that simultaneously safeguards patient privacy, reduces edge‑side energy consumption, and enhances subgroup‑aware safety in dermatological AI. By coupling uncertainty‑driven computation with interpretable metadata‑derived risk, it offers a compelling blueprint for trustworthy, sustainable medical inference in resource‑constrained edge environments.