利用機器學習進行皮膚神經形態定量分析：糖尿病神經病變的診斷與臨床意義

Abstract

目的： 本研究旨在：(1) 運用機器學習演算法建立並驗證新型表皮內神經纖維（IENF）生物標記，以協助診斷糖尿病患者之小纖維神經病變（SFN）；(2) 評估此類新型指標的臨床應用價值與診斷準確度。 方法： 本研究納入糖尿病神經病變患者與對照組受試者，導入機器學習自動化分析技術進行定量，進而發展出以面積為基礎的 IENF 形態測量（IENFa）參數。診斷效益透過受試者操作特徵曲線（ROC）進行評估；另結合代謝數據與電生理檢驗結果之相關性分析，以剖析各 IENFa 指標的臨床相關性。 結果： 糖尿病神經病變組（n=48）與對照組（n=63）之年齡及性別比例相近。在對照組中，IENFa 參數隨年齡增加而呈現遞減趨勢，其中僅 IENFd 與 IENFa/A 展現性別差異。基於 (1) 與 IENFd 的相關性分析，以及 (2) 以 IENFd 判定 SFN 之 ROC 診斷效益（AUC：0.91–0.95，P>0.05），各項 IENFa 參數均展現出不相上下的診斷準確度。此外，IENFa 標記與腓腸神經感覺神經動作電位（SNAP）波幅呈顯著正相關。 結論： 自動化 IENFa 定量技術具備高度省時特性，於糖尿病 SFN 的診斷效能上不僅不亞於傳統 IENFd，且具備優異的再現性與信度。再者，IENFa 指標亦能同步反映糖尿病神經病變中伴隨的大神經纖維病變。鑑於 IENFa 反映的是 IENF 的總累計面積，此發現亦提示糖尿病神經病變過程可能伴隨整體性的軸突萎縮變化。

This study sought to establish and validate novel area-based intraepidermal nerve fiber (IENF) biomarkers using a machine learning–assisted approach for the diagnosis of diabetic small fiber neuropathy (SFN), and to investigate their diagnostic utility and clinical relevance. Individuals with diabetic neuropathy and healthy controls were enrolled. A machine learning–based automated image analysis system was developed to quantify area-based IENF (IENFa) parameters. Their performance of the diagnosis was investigated by receiver operating characteristic (ROC) analysis, while their clinical significance was examined by assessing associations with metabolic characteristics and electrophysiological findings. A total of 48 patients with diabetic neuropathy and 63 control subjects were included, with the two groups being comparable in terms of age and sex distribution. In the control cohort, all IENFa parameters declined with increasing age, whereas sex-related differences were observed only for conventional IENF density (IENFd) and the IENFa/A parameter. The newly developed IENFa biomarkers showed strong agreement with IENFd and demonstrated comparable accuracy for identifying IENFd-defined SFN, with areas under the ROC curve ranging from 0.91 to 0.95 and no significant differences among the parameters (P > 0.05). In addition, the IENFa measurements were associated with sensory nerve action potential amplitudes of sural nerve with significance, indicating a close relationship with large-fiber dysfunction. In conclusion, machine learning–based automated quantification of IENFa provides a rapid, reliable, and objective method for diagnosing diabetic SFN, with diagnostic performance comparable to that of conventional IENFd. Because IENFa reflects the cumulative area of all intraepidermal nerve fibers, these findings further suggest that diabetic neuropathy is characterized by diffuse axonal atrophy involving both small and large peripheral nerve fibers.