血浆ATR-FTIR光谱结合机器学习用于鼻咽癌分类。

文献信息

中文摘要

背景： 鼻咽癌（NPC）常在晚期才被诊断，因此需要微创方法来支持检测。本研究评估了血浆衰减全反射傅里叶变换红外（ATR-FTIR）光谱结合机器学习，以区分NPC与临床健康个体。
方法： 对51例经组织学确诊的NPC病例和51例年龄、性别匹配的临床健康对照者的血浆样本在4000-600 cm-1范围内进行分析。使用Mann-Whitney U检验评估21个光谱峰处的差异。使用全光谱、指纹区、选定峰和低秩光谱表示，通过重复交叉验证和探索性年龄、性别分层分析，评估了七种机器学习算法。
结果： 13个峰在中位吸光度上显示出显著的组间差异（p < 0.01）。尽管原始光谱分布存在大量重叠，但在完整队列中，当使用选定峰和低秩光谱表示时，模型性能有所提高：最佳情况AUC分别为全光谱0.6303 ± 0.0384、指纹区0.6330 ± 0.0333、选定峰0.7066 ± 0.0452、低秩表示0.7404 ± 0.0422。神经网络模型使用低秩特征取得了最佳总体性能，准确率（ACC）为0.7115 ± 0.0456。探索性亚组估计在年龄和性别定义的亚组间存在差异。然而，由于亚组样本量有限且性能估计存在变异性，对解释和普遍性需谨慎。
结论： 血浆ATR-FTIR光谱结合机器学习在NPC与临床健康对照者之间显示出中等程度的区分能力，使用低秩光谱特征和前馈神经网络获得最高性能。需要更大规模的独立队列进行验证。

英文摘要

BACKGROUND: Nasopharyngeal carcinoma (NPC) is often diagnosed at advanced stages, creating a need for minimally invasive approaches to support detection. This study evaluated plasma attenuated total reflectance Fourier transform infrared (ATR-FTIR) spectroscopy, combined with machine learning, to distinguish NPC from clinically healthy individuals.
METHODS: Plasma samples from 51 histologically confirmed NPC cases and 51 age- and sex-matched clinically healthy controls were analyzed across 4000-600 cm-1. Differences at 21 spectral peaks were assessed using the Mann-Whitney U test. Seven machine-learning algorithms were evaluated using the full spectrum, fingerprint region, selected peaks, and low-rank spectral representations with repeated cross-validation and exploratory age- and sex-stratified analyses.
RESULTS: Thirteen peaks showed significant between-group differences in median absorbance (p < 0.01). Despite substantial overlap in the original spectral distributions, model performance in the complete cohort increased when selected peaks and low-rank spectral representations were used: the best case AUCs were 0.6303 ± 0.0384 for the full spectrum, 0.6330 ± 0.0333 for the fingerprint region, 0.7066 ± 0.0452 for selected peaks, and 0.7404 ± 0.0422 for the low-rank representation. The neural network model achieved the best overall performance using low-rank features, with an accuracy (ACC) of 0.7115 ± 0.0456. Exploratory subgroup estimates varied across age- and sex-defined subsets. However, due to limited subgroup sizes and variability in the performance estimates, interpretation and generalizability was cautioned.
CONCLUSION: Plasma ATR-FTIR spectroscopy combined with machine learning showed moderate discrimination between NPC and clinically healthy controls, with the highest performance obtained using low-rank spectral features and a feedforward neural network. Larger independent cohorts are needed for validation.