周邊動脈疾病(PAD)並非侷限於下肢之局部血管病變,而是全身性動脈粥狀硬化之周邊表現,常與慢性腎臟病(CKD)及心臟疾病高度併發。由於PAD 早期症狀具隱匿性,且現行踝臂血壓比值(ABI)易受血管鈣化影響,於PAD 合併 CKD 等共病族群中準確度顯著下降。此外,現行血管與心臟評估工具多獨立運作,缺乏同步整合之雙重篩檢機制,使心血管共病之潛在風險易被忽略。

 

本研究基於既有之四肢同步生理訊號擷取架構與訊號品質篩選機制(SSSM),建置一套整合周邊血管評估與心臟電生理篩檢之 AI雙路徑評估系統。於血管路徑中,經 SSSM 品質篩選後之光電容積脈搏波(PPG)訊號,除擷取脈波傳導時間(PTT)外,進一步量化脈波上升時間、上升斜率與下降斜率等多維時域波形特徵,並採用極限梯度提升 (XGBoost)演算法建立健康族群、PAD患者與PAD 合併 CKD患者之三分類診斷模型。於心臟路徑中,單導程心電圖(ECG)訊號不經SSSM篩選,直接輸入以xresnet1 d101 架構為基礎、經單導程遷移學習調適之心臟深度學習診斷模型,評估正常(NORM)、傳導障礙(CD)、ST/T波變化(STTC)、心肌梗塞(MI)與心室肥大(HYP)五類心臟病變風險。 兩模型獨立訓練、互不干涉,於推論完成後再進行跨模態關聯分析,探討周邊血管嚴重度分級與心臟病變風險是否呈現具臨床意義之一致趨勢。

周邊動脈疾病(PAD)並非侷限於下肢之局部血管病變,而是全身性動脈粥狀硬化之周邊表現,常與慢性腎臟病(CKD)及心臟疾病高度併發。由於PAD 早期症狀具隱匿性,且現行踝臂血壓比值(ABI)易受血管鈣化影響,於PAD 合併 CKD 等共病族群中準確度顯著下降。此外,現行血管與心臟評估工具多獨立運作,缺乏同步整合之雙重篩檢機制,使心血管共病之潛在風險易被忽略。

本研究基於既有之四肢同步生理訊號擷取架構與訊號品質篩選機制(SSSM),建置一套整合周邊血管評估與心臟電生理篩檢之 AI雙路徑評估系統。於血管路徑中,經 SSSM 品質篩選後之光電容積脈搏波(PPG)訊號,除擷取脈波傳導時間(PTT)外,進一步量化脈波上升時間、上升斜率與下降斜率等多維時域波形特徵,並採用極限梯度提升 (XGBoost)演算法建立健康族群、PAD患者與PAD 合併 CKD患者之三分類診斷模型。於心臟路徑中,單導程心電圖(ECG)訊號不經SSSM篩選,直接輸入以xresnet1 d101 架構為基礎、經單導程遷移學習調適之心臟深度學習診斷模型,評估正常(NORM)、傳導障礙(CD)、ST/T波變化(STTC)、心肌梗塞(MI)與心室肥大(HYP)五類心臟病變風險。 兩模型獨立訓練、互不干涉,於推論完成後再進行跨模態關聯分析,探討周邊血管嚴重度分級與心臟病變風險是否呈現具臨床意義之一致趨勢。

實驗結果顯示,相較於第一代系統僅採用單-PTT指標,本研究之多維PPG特徵血管模型將三分類準確度由81.97%提升至91.80%,巨集平均 Fl-score由80.27%提升至 88.81%;其中臨床上易受血管鈣化干擾之PAD 合併 CKD 族群,其敏感度亦由57.14% 提升至81.25%。心臟診斷模型方面,經單導程遷移學習後之三種子集成模型於PTB-XL 測試集達巨集平均 AUC 0.8389(95% CI [0.8279, 0.8493]),較未調適之零樣本模型 (0.7115)顯著提升,其中心肌梗塞類別提升幅度最大(+0.2572)。跨模態關聯分析顯示,NORM、CD 與STTC 類別之陽性率隨血管疾病嚴重度呈單調變化趨勢,心肌梗塞陽性率亦由0%、14.3%上升至83.3%;進一步以Spearman 與 Kendall 等級相關係數進行統計檢定,五類心臟病變之陽性機率與血管嚴重度分級間均呈顯著相關,為周邊血管病變嚴重度與心臟病理負荷之協同變化關係提供統計上的支持證據。本系統成功將評估維度由單一血管疾病篩檢擴展至周邊血管與心臟之雙重自動化評估,具備不依賴血管可壓縮性之非侵入式篩檢潛力與高臨床應用價值。

 

Peripheral Artery Disease (PAD) is not merely a localized vascular lesion confined to the lower extremities, but a peripheral manifestation of systemic atherosclerosis that frequently co-occurs with chronic kidney disease (CKD) and cardiovascular comorbidities. Because early-stage PAD is often clinically silent, and the Ankle-Brachial Index (ABI)-the current standard-of-care screening tool-suffers from significant diagnostic failure in cohorts with severe vascular calcification such as PAD+CKD patients, an alternative screening approach independent of vascular compressibility is urgently needed. Moreover, existing peripheral vascular and cardiac assessment tools are predominantly managed as independent clinical workflows, leaving the potential risk of concurrent cardiovascular comorbidity largely unaddressed.

Building upon the existing four-limb synchronous physiological signal acquisition infrastructure and the Simultaneous Signal Selection Model (SSSM), this study develops an AI evaluation architecture that integrates peripheral vascular assessment with cardiac electrophysiological screening through two independently operating pathways. In the vascular pathway, photoplethysmography (PPG) signals that pass SSSM quality screening are used to extract, beyond Pulse Transit Time (PTT), three additional multi-dimensional time-domain geometric waveform features-Rise Time, Upstroke Slope, and Downstroke Slope-which are fed into an Extreme Gradient Boosting (XGBoost) classifier to perform three-class discrimination among Normal, PAD, and PAD+CKD subjects. In the cardiac pathway, the single-lead electrocardiography (ECG) signal bypasses SSSM screening entirely and is directly input into a cardiac diagnostic model built on the xresnet1 d101 architecture, adapted via single-lead transfer learning, to evaluate five diagnostic classes: Normal (NORM), Conduction Disturbance (CD), ST/T Change (STTC), Myocardial Infarction (MI), and Hypertrophy (HYP). The two models are trained and executed entirely independently, without feature-level fusion; a cross-modal correlation analysis is subsequently performed on their outputs to examine whether peripheral vascular severity classification exhibits a clinically consistent trend with cardiac diagnostic risk.

Experimental results show that, compared with the first-generation system relying solely on a single PTT metric, the proposed multi-dimensional PPG feature vascular inference model raises three-class classification accuracy from 81.97% to 91.80% and macro-averaged Fl-score from 80.27% to 88.81%; sensitivity for the PAD+CKD class-which is particularly susceptible to interference from vascular calcification also improves from 57.14% to 81.25%. For the cardiac diagnostic model, the three-seed probability-averaged ensemble obtained via single-lead transfer learning achieves a Macro-AUC of 0.8389 (95% bootstrap CI [0.8279, 0.8493]) on the PTB-XL Lead-I test set, a marked improvement over the zero-shot channel-sliced baseline (0.7115), with the largest gain observed in the Myocardial Infarction class (+0.2572). The cross-modal correlation analysis further reveals that the positive rates of the NORM, CD, and STTC classes change monotonically with increasing vascular disease severity, while the MI positive rate also rises from 0% to 14.3% to 83.3% across the Normal, PAD, and PAD+CKD groups; this monotonic association was further confirmed through rank-based correlation analysis, which showed statistically significant Spearman's and Kendall's correlation coefficients between vascular severity and cardiac model output across all five diagnostic classes, even after Bonferroni correction for multiple comparisons. These findings suggest a coherent co-variation between peripheral vascular severity and cardiac pathological burden. Overall, the proposed system successfully extends the evaluation scope from single-disease vascular screening to a dual-pathway automated assessment of both peripheral vascular and cardiac status, offering a non-invasive screening approach independent of vascular compressibility and demonstrating strong clinical translational value.

 

本研究基於既有之四肢同步生理訊號擷取架構與訊號品質篩選機制(SSSM),建置一套整合周邊血管評估與心臟電生理篩檢之 AI雙路徑評估系統。於血管路徑中,經 SSSM 品質篩選後之光電容積脈搏波(PPG)訊號,除擷取脈波傳導時間(PTT)外,進一步量化脈波上升時間、上升斜率與下降斜率等多維時域波形特徵,並採用極限梯度提升 (XGBoost)演算法建立健康族群、PAD患者與PAD 合併 CKD患者之三分類診斷模型。於心臟路徑中,單導程心電圖(ECG)訊號不經SSSM篩選,直接輸入以xresnet1 d101 架構為基礎、經單導程遷移學習調適之心臟深度學習診斷模型,評估正常(NORM)、傳導障礙(CD)、ST/T波變化(STTC)、心肌梗塞(MI)與心室肥大(HYP)五類心臟病變風險。 兩模型獨立訓練、互不干涉,於推論完成後再進行跨模態關聯分析,探討周邊血管嚴重度分級與心臟病變風險是否呈現具臨床意義之一致趨勢。

 

實驗結果顯示,相較於第一代系統僅採用單-PTT指標,本研究之多維PPG特徵血管模型將三分類準確度由81.97%提升至91.80%,巨集平均 Fl-score由80.27%提升至 88.81%;其中臨床上易受血管鈣化干擾之PAD 合併 CKD 族群,其敏感度亦由57.14% 提升至81.25%。心臟診斷模型方面,經單導程遷移學習後之三種子集成模型於PTB-XL 測試集達巨集平均 AUC 0.8389(95% CI [0.8279, 0.8493]),較未調適之零樣本模型 (0.7115)顯著提升,其中心肌梗塞類別提升幅度最大(+0.2572)。跨模態關聯分析顯示,NORM、CD 與STTC 類別之陽性率隨血管疾病嚴重度呈單調變化趨勢,心肌梗塞陽性率亦由0%、14.3%上升至83.3%;進一步以Spearman 與 Kendall 等級相關係數進行統計檢定,五類心臟病變之陽性機率與血管嚴重度分級間均呈顯著相關,為周邊血管病變嚴重度與心臟病理負荷之協同變化關係提供統計上的支持證據。本系統成功將評估維度由單一血管疾病篩檢擴展至周邊血管與心臟之雙重自動化評估,具備不依賴血管可壓縮性之非侵入式篩檢潛力與高臨床應用價值。