AeroSpectra Sentinel: LLM Prompt-Chaining Hits 91% Asthma Screening Accuracy

AeroSpectra Sentinel, developed by Aueaphum Aueawatthanaphisut, is an auditable decision-support workflow that addresses acute asthma risk assessment by combining short-time Fourier transform (STFT) respiratory sound analysis, lightweight machine learning screening, clinical feature fusion, and a five-stage large language model (LLM) prompt-chaining process. The system separates signal acquisition, preprocessing, acoustic feature extraction, ML screening, clinical guardrails, and FHIR-ready reporting into distinct stages. Evaluated on a public dataset of 1,211 WAV recordings (584 stratified subset for binary classification), a random forest achieved 91.10% accuracy and 78.69% F1-score for asthma vs. non-asthma screening. A feature-based multilayer perceptron reached 89.73% accuracy and 78.26% F1, while a compact log-spectrogram CNN managed 73.29% accuracy and 55.17% F1. Multiclass classification across five labels hit 77.40% accuracy and 77.23% macro-F1.

The LLM workflow was tested on 40 simulated clinical vignettes comparing one-shot prompting, prompt chaining, prompt chaining with guardrails, and prompt chaining with guardrails plus FHIR schema validation. The guardrail-plus-schema variant delivered the strongest simulated safety and documentation consistency. AeroSpectra Sentinel is explicitly a research prototype—not a diagnostic device—intended to demonstrate transparent reasoning and safe escalation logic for acute asthma risk assessment. The approach highlights how LLM prompt-chaining can add auditability and clinical context beyond conventional audio-only classifiers.