AI-Based Spectral Analysis for Bacterial Biomarker Detection in Wound Diagnostics

Abstract

The management of chronic wound infections presents a significant clinical challenge, often exacerbated by diagnostic delays and the limitations of subjective clinical assessment [1]. While point-of-care fluorescence imaging systems offer non-invasive visualization, precise biochemical quantification remains severely obstructed by complex, non-linear optical phenomena such as the Inner Filter Effect (IFE). Consequently, translating high-fidelity spectroscopy into cost-effective, edge-deployable diagnostic devices without a catastrophic loss of classification accuracy remains a major engineering bottleneck. Here, we establish a dual-track roadmap for clinical device engineering by systematically evaluating optical biomarker diagnostics for Coproporphyrin I (CpI) and Protoporphyrin IX (PpIX) across two distinct hardware paradigms: a continuous high-resolution miniature spectrometer and a constrained, discrete low-resolution photodiode sensor. For the high-resolution data streams, Convolutional Neural Networks (CNNs) and Spectral Transformers demonstrated a robust capacity to resolve non-linear optical variations, identifying optimal classification boundaries within controlled experimental datasets and serving as a scalable framework for future in vivo integration. Conversely, for the highly constrained low resolution sensor array, the research isolated a physical indistinguishability limit stemming from coarse spectral resolution. This barrier was systematically overcome through domain-aware feature engineering and the realignment of diagnostic targets into a pragmatic 5×5 triage matrix. This physical-feature transformation enabled a minimalist Decision Tree architecture to achieve diagnostic parity, yielding a 98% classification accuracy. Concurrently, hardware profiling of the structurally pruned low-resolution convolutional models validated the embedded “race to sleep” paradigm [2]; the quantized 8-bit Integer (INT8) micro variant of the low resolution model executed edge inference in just 0.85ms while consuming an ultra-low 0.17mJ of energy on an ESP32-S3 microcontroller. Ultimately, this physics-first approach provides a transparent, certifiable foundation for immediate low-resolution edge deployment, while the deep-learning frameworks serve as a foundational proof-of-concept, demonstrating the diagnostic potential of these algorithms once comprehensive real-world datasets become available.