Scholarlink

<p>Medical Internet of Things (MIoT) systems enable continuous cardiac monitoring, but practical deployment is limited by three issues: heavy computation at the edge, limited interpretability, and vulnerability to cyber-attacks that can corrupt signals and degrade inference. We propose CLARITY-AI 2.0, a lightweight and trustworthy arrhythmia detection framework for MIoT that combines efficient ECG feature extraction, interpretable prediction, and security-aware trust assessment. The model performs arrhythmia inference using a compact feature-based learner and generates clinician-facing explanations via SHAP attributions, which are converted into structured natural-language reports using an LLM. To improve trust under hostile network conditions, an intrusion detection module outputs an attack probability and a trust score that flags unreliable inputs. On the MIT-BIH benchmark, CLARITY-AI 2.0 achieves an F1 score of 0.928 and an AUC of 0.985 for anomaly detection. It also generalizes under dataset shift with an AUC of 0.940 on PTB-XL and an AUC of 0.915 on Chapman (zero-shot evaluation). For edge feasibility, deployment on an ESP32 reduces the footprint to 350 KB, peak RAM to 120 KB, and inference latency to 8.1 ms, supporting real-time operation on resource-constrained devices. These results indicate that CLARITY-AI 2.0 is an interpretable, efficient, and security-aware approach toward scalable MIoT-based cardiac monitoring.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Lightweight and interpretable edge intelligence AI with intrusion detection for trustworthy cardiac arrhythmia in medical IoT

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月24日

查看全文
  • Muhammad Imran Khalid,
  • Altaf Hussain,
  • Nasir Hussain,
  • Tamim Alkhalifah

摘要

Medical Internet of Things (MIoT) systems enable continuous cardiac monitoring, but practical deployment is limited by three issues: heavy computation at the edge, limited interpretability, and vulnerability to cyber-attacks that can corrupt signals and degrade inference. We propose CLARITY-AI 2.0, a lightweight and trustworthy arrhythmia detection framework for MIoT that combines efficient ECG feature extraction, interpretable prediction, and security-aware trust assessment. The model performs arrhythmia inference using a compact feature-based learner and generates clinician-facing explanations via SHAP attributions, which are converted into structured natural-language reports using an LLM. To improve trust under hostile network conditions, an intrusion detection module outputs an attack probability and a trust score that flags unreliable inputs. On the MIT-BIH benchmark, CLARITY-AI 2.0 achieves an F1 score of 0.928 and an AUC of 0.985 for anomaly detection. It also generalizes under dataset shift with an AUC of 0.940 on PTB-XL and an AUC of 0.915 on Chapman (zero-shot evaluation). For edge feasibility, deployment on an ESP32 reduces the footprint to 350 KB, peak RAM to 120 KB, and inference latency to 8.1 ms, supporting real-time operation on resource-constrained devices. These results indicate that CLARITY-AI 2.0 is an interpretable, efficient, and security-aware approach toward scalable MIoT-based cardiac monitoring.