[中图分类号] R737.33;R730.4 [文献标志码] A

Establishment and assessment an early screening model for cervical cancer based on single-cell Raman spectroscopy combined with machine learning algorithmsMA Dongmei， ZHAO Wenjie， LIU Shihai ， XU Haicang，CAI Duo，JI Yuetong，XU Jian，GUO Cancan，MA Bo，PAN Huazheng（ ， University， 266oo3，China）

[ABSTRACT]ObjectiveTo establish an early screning model forcervical cancer based on single-cellRaman spectroscopy （SCRS）combined with machine learningalgorithms，and toasess theperformance themodel.MethodsCervical exfoliated cellsamples werecolectedfrom128patients who weretreatedinourhospitalfromSeptember2023 toJune2024，among whom65 had normalresultsThinPrepcytologic test（TCT），35hadabnormal TCTresults，and 28did notreceiveTCT.Rlanguage was used to divide the 10O cervical exfoliated cellsamples with TCT results into training set and test set at a ratio 8：2 ，and SCRS was performedforallsamples.BasedontheSCRSdata therainingset，7 machinelearning algorithms（KNN，PLS，LDA，RF， SVM，SVMRBF，and Stack）wereused toestablishanearlyscrening modelforcervicalcancer，whichwasappliedinthe testset toidentifytheoptimalmodel.optimalmodelwasthenusedtopredicttheTCTresults1Ocervicalexfoliatedcellsamples in thetraining and testsets，whichwerecomparedwiththeactualTCTresults.remaining 28 samples without priorTCTresults wereusedasavalidationsetand weresubjectedtoTCT，nd theoptimalmodelwasusedtopredict theTCTresultsthesesamples，whichwerecompared withtheactualTCTresults.Resultsre weresignificantdiferencesintherelativeintensities characteristic Raman peaks at 874，935，1 024，1 119，1 250，1 328，1 569，and 1642cm^-1 between the cervical exfoliated cells negativeforintraepitheliallesionormalignancy，atypicalsquamouscellsundeterminedsignificane，andthecervicalexfoliated celslow-gradesquamous intraepitheliallesion.Amongthe7algorithms，thestacking modelshowedthebestperformance，with anAUC O.987，anaccuracy 99.2% ，asensitivity 98.9% ，and a specificity 99.3% .In both training and test sets，the results predicted bythe Stack model were relativelyhighly consistentwithactual TCT results， with an accuracy 91.0% ，a sensitivity 91.0% ，a specificity 87.4% ，and an Fl-score 90.3% .In the validation set，the Stack model achieved an accuracy 96.4% ，a sensitivity 100‰ ，a specificity 95.5% ，and an F1-score 92.3% in predicting TCT results.Conclusion early screening modelforcervicalcancerbasedonSCRSand machinelearning algorithms hasagoodperformanceandcanbeusedasa noninvasive，efficient，and rapid tool to facilitate the early screening cervical cancer.

[KEY WORDs]Uterine cervical neoplasms；Single-cell analysis；Spectrum analysis，Raman；Machine learning；Algo rithms;Early detection cancer

宫颈癌是生殖系统常见肿瘤，在全球女性罹患的肿瘤中居第二位，其发病率和病死率逐年上升，严重威胁女性健康[1]。目前宫颈癌的临床筛查主要依赖液基薄层细胞学检测（TCT）以及人乳头瘤病毒（HPV）检测，确诊仍需依靠宫颈活检，这一过程存在假阳性率高、主观性强、有创且检测周期较长等局限性。因此目前临床亟需探索一种快速非侵入性的宫颈癌早期精准筛查方法。拉曼光谱检测是一种基于光与物质相互作用的无损分析技术，通过测量分子散射光的频移信息识别样本的化学组成和分子结构。拉曼光谱检测过程无需染色或破坏细胞结构，可保留样本完整状态，便于后续复检或多组学联合分析，在肿瘤精准诊断、神经系统疾病筛查及病原体鉴别等领域具有无损、快速等优势[3]。拉曼光谱对核酸、蛋白质二级结构及脂质氧化等生物大分子异常信号的捕捉具有极高的灵敏度，能够反映细胞从正常向癌变的早期转变过程[4]。而宫颈癌在早期病变阶段，宫颈脱落细胞核内DNA含量增加，磷酸盐骨架的拉曼峰增强，同时蛋白质中酪氨酸和色氨酸的拉曼信号发生明显漂移[5]。目前已有研究将单细胞拉曼光谱（SCRS）检测结合机器学习算法用于肝癌、肾癌以及宫颈癌等患者的血液及病理组织样本检测，以实现肿瘤的早期筛查。本研究利用微流控芯片采集宫颈脱落细胞样本的SCRS，构建基于SCRS和机器学习算法的TCT结果预测模型，尝试提出一种高效、无创、精准的宫颈癌早期筛查方法，现将结果报告如下。……