Accuracy of Raman spectroscopy in laryngeal carcinoma diagnosis: a meta-analysis

Abstract

Background

This study explores the accuracy of Raman spectroscopy in diagnosing laryngeal carcinoma.

Methods

A comprehensive search was performed in PubMed, Web of Science, and Embase (retrieval deadline: March 2025) to locate studies that presented data on true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN) regarding the use of Raman spectroscopy for diagnosing laryngeal carcinoma. Using either random or fixed-effects models, we computed pooled sensitivity, specificity, diagnostic odds ratio (DOR), and the area under the summary receiver operating characteristic (SROC) curve. The methodological quality was evaluated with QUADAS-2, and Deek’s test was employed to assess publication bias.

Results

Seven studies involving 1076 samples in total were analyzed. The overall sensitivity and specificity were determined to be 0.88 (95% confidence interval [CI] 0.86–0.90) and 0.91 (95% CI 0.88–0.93), respectively. Positive likelihood ratios (PLR) and negative likelihood ratios (NLR) were calculated at 9.70 (95% CI 7.13–13.21) and 0.13 (95% CI 0.11–0.16), respectively, resulting in a diagnostic odds ratio (DOR) of 74.07 (95% CI 39.0–140.9). The summary receiver operating characteristic (SROC) curve demonstrated an area under the curve (AUC) of 0.95 (95% CI 0.94–0.98). Furthermore, subgroup analysis revealed significant improvements in specificity over time (p = 0.02), and no significant regional or algorithmic heterogeneity was found.

Conclusion

Raman spectroscopy exhibits remarkable diagnostic accuracy in detecting laryngeal carcinoma, with a pooled sensitivity of 88% and a specificity of 91%. This highlights its promise as a non-invasive supplement for diagnosis.

Supplementary Information

The online version contains supplementary material available at 10.1186/s40001-025-03309-5.

Introduction

Laryngeal carcinoma ranks among the most common types of malignant tumors found in the head and neck area. According to estimates from the American Cancer Society, roughly 13,000 new instances of laryngeal carcinoma are diagnosed annually, leading to about 3700 deaths each year. Early detection and diagnosis, coupled with effective treatment, can significantly enhance the 5-year survival rate [1]. Currently, histological examination serves as the gold standard for tissue diagnosis. The typical approach for intraoperative diagnosis and margin evaluation is frozen section analysis; nonetheless, this method is expensive and requires considerable time, as postoperative pathology outcomes frequently take several hours or even days to obtain. Furthermore, only a limited number of critical margins are extracted for testing, which increases the risk of missed diagnoses [2]. Consequently, there is a pressing need for more convenient diagnostic tools that offer higher specificity and sensitivity in the diagnosis of laryngeal carcinoma.

Raman spectroscopy, a non-invasive analytical technique based on molecular vibrational fingerprints, can accurately elucidate the structural differences in biomolecular structures by capturing characteristic frequency shifts resulting from the interaction between photons and molecules. This method is particularly effective in tumor diagnosis, as it identifies specific protein conformational changes and lipid metabolic abnormalities in cancerous tissues, achieving high specificity in differentiating between breast carcinoma, gastric carcinoma, and normal tissues [3]. Compared to traditional biopsy, this technology offers three core advantages: non-destructive detection, real-time feedback, and molecular resolution, making it particularly suitable for rapid intraoperative diagnostic scenarios. Recent advancements in surface-enhanced Raman spectroscopy (SERS), tip-enhanced Raman spectroscopy (TERS), and shell-separated nanoparticle enhanced Raman spectroscopy (SHINERS) have improved detection sensitivity to the single-molecule level, significantly enhancing the early detection rate of microscopic lesions [4]. Moreover, this technology exhibits excellent sample adaptability, ranging from in situ detection of live cells to the analysis of complex biological fluids. Its spectral imaging capability allows for the simultaneous acquisition of information on tissue spatial heterogeneity, providing crucial molecular imaging evidence for clinical precision diagnosis and treatment [5]. Although single-center studies have reported on the diagnostic capabilities of Raman spectroscopy for laryngeal carcinoma, differences in findings could stem from inconsistencies in experimental designs, methods of sample collection, and the choice of Raman spectroscopy techniques. For the first time, we strictly followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines [4]. We performed a systematic review combined with meta-analysis to assess the diagnostic effectiveness of Raman spectroscopy in the context of laryngeal carcinoma, with the goal of elucidating its efficacy in confirming the presence of this type of carcinoma.

Method

Literature review

We invited two attending physicians specializing in otolaryngology to conduct a comprehensive search of the database content, with a deadline extending to March 2025. The databases utilized for this search included PubMed, Embase, and Web of Science. We employed medical subject headings such as “larynx”, “cancer”, and various iterations of “Raman spectroscopy” to facilitate our search. Furthermore, we performed a supplementary review of the references from the retrieved articles to ensure the thoroughness of the literature extracted. In the event of any disputes regarding inclusion, the deputy chief physician would render the final decision.

Include inclusion criteria and exclusion criteria

We systematically and meticulously screened all relevant articles based on the following inclusion criteria: (1) Raman spectroscopy utilized as an independent diagnostic tool for distinguishing cancerous tissue from normal tissue; (2) the selected literature must consist of human tissues or specimens; (3) relevant data must be extractable; (4) all articles must be published in English. The exclusion criteria were as follows: (1) trials that do not involve human specimens were excluded; (2) articles that do not conform to the specified types, such as conference abstracts, letters, reviews, and case reports, were excluded; (3) articles from which relevant data cannot be accurately extracted were excluded; (4) non-English articles were excluded; (5) articles that do not meet the requirements after a final review by associate chief physicians were excluded.

Data fetch

In this study, we included essential information such as the title, year, country, sample type, type of Raman spectrum, true positive (TP), false positive (FP), true negative (TN), false negative (FN), and other relevant data. Ultimately, we compiled these findings into a comprehensive table. Notably, the two tables concerning attending physicians exhibited a high degree of consistency.

Quality assessment and publication bias

We conducted a quality assessment of the seven included studies using R for Windows, based on the Diagnostic Accuracy Research Quality Assessment (QUADAS-2) checklist. The Stata software version 14.0 (STATA Corp, TX, USA) was utilized, along with the Deeks funnel plot asymmetry test; an asymmetrical funnel shape or a p-value of less than 0.05 indicates the presence of publication bias.

Statistical analysis

To begin with, it is essential to extract the relevant data pertaining to true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN). Once this data is gathered, we can proceed to calculate several important statistical metrics that inform our understanding of the diagnostic tool’s performance. These metrics include summary sensitivity, specificity, accuracy, positive and negative likelihood ratios (LRs), as well as the diagnostic odds ratio (DOR). Additionally, we can assess the summary receiver operating characteristic (SROC) curve, the area under the curve (AUC), and the 95% confidence intervals (CI). For the statistical analysis, we employ the software Stata version 14.0, which provides robust capabilities for such tasks. The interpretation of the AUC value is particularly significant in evaluating the diagnostic tool. When the AUC exceeds 0.9, it signifies that the diagnostic tool is operating at an excellent level of accuracy. Conversely, if the AUC falls between 0.8 and 0.9, the diagnostic performance is deemed good. An AUC in the range of 0.7–0.8 represents a moderate level of diagnostic effectiveness, while a value below 0.7 is regarded as poor, indicating a need for improvement in the diagnostic method or tool.

Results

Literature search and study selection

A comprehensive search was conducted in PubMed, EMBASE and Web of Science databases, and a total of 517 studies were identified. The specific search strategy is detailed in supplementary document 1. Following this initial identification, we proceeded to eliminate any duplicate entries and conducted a thorough screening of the titles and abstracts to assess their relevance. From this rigorous evaluation process, we ultimately selected 10 studies that fulfilled the criteria necessary for a full-text review. After a detailed examination of these texts, we determined that 7 of the studies qualified for inclusion in the subsequent meta-analysis. A flowchart summarizing this selection process is provided below for clarity Fig. 1.

Fig. 1.

Flowchart showing search strategy

Characteristics, quality assessment and publication bias

The overall characteristics of the studies included in this review are detailed in Table 1.

Table 1.

Study characteristics

Reference	Year	Nation	Age	N1	N2	N3	Sample type	Diagnostic algorithm	Excitation frequency (cm−1)	Data preprocessing	Study design
Stone [6]	2000	United Kingdom	53	76	52	24	Ex vivo tissue	PLS-DA	400—1800	A first-order Savitsky–Golay filter	Diagnostic feasibility study
Lau [7]	2005	Canada	55	31	18	13	Ex vivo tissue	PCA	950—1650	NR	Diagnosticcohort study with cross-sectional features
The [8]	2009	Singapore	NR	50	20	30	Ex vivo tissue	PCA-LDA	800—1800	A first-order Savitsky–Golay filter	Diagnostic study
Lin [9]	2012	Singapore	NR	94	22	72	In vivo tissue	PCA-LDA	2800–3020	A first-order Savitsky–Golay filter	Diagnostic cohort study
Lin [10]	2016	Singapore	NR	101	71	30	In vivo tissue	PLS-DA	400 − 3600	A first-order Savitsky–Golay filter	Diagnostic cohort study
Li [11]	2022	USA	NR	707	207	500	Ex vivo tissue	PCA	509 − 3,978	Asymmetric least squares smoothing, Savitzky–Golay filter	Diagnostic cohort study
Van [12]	2023	Netherlands	57	17	0	17	Ex vivo tissue	PCA-LDA	2500 − 4000	A first-order Savitsky–Golay filter	Diagnostic cohort study

USA United States of America, PLS-DA partial least squares discriminant analysis, N1 total sample size, N2 total number of healthy controls or adjacent samples, N3 total number of cancer samples, PCA principal component analysis, PCA-LDA principle component analysis integrated with linear discriminant analysis

Among these studies, seven were published in the English language. Altogether, these investigations analyzed a comprehensive total of 1076 samples. However, McGregor study was excluded due to small sample size (only a small number of data from 430 patients) and lack of raw spectral data. This exclusion highlights the need for careful consideration of participant selection criteria when interpreting the findings of the included research. Among these studies, three employed PCA-LDA for spectral analysis, while two utilized PCA and another two implemented PLS-DA. Histopathology served as the gold standard across all studies. The findings from the quality assessment are illustrated in the QUADAS-2 plot presented in Fig. 2. This plot shows that a significant number of the studies met the criteria outlined in the QUADAS-2 guidelines, suggesting a generally high standard of methodological rigor across the assessed research. Furthermore, an evaluation of publication bias was conducted using the Deeks funnel plot asymmetry test. The results of this analysis indicated no notable publication bias, as evidenced by a bias value of - 0.14 and a P-value of 0.12, which is shown in Fig. 3. This supports the reliability of the studies included in this assessment.

Fig. 2.

A Risk of bias and applicability concerns evaluation of included studies in pool. B Risk of bias and applicability concerns evaluation of included studies individually

Fig. 3.

The Deek’s funnel plot asymmetry test indicated that there was no publication bias (bias =  − 0.14, P = 0.12)

The diagnostic potential of Raman spectroscopy in the context of laryngeal carcinoma has been evaluated through a review of seven studies, revealing a sensitivity range between 0.69 and 0.93. When these results were combined, the overall pooled sensitivity was determined to be 0.88, with a 95% confidence interval (CI) of 0.86 to 0.90. Such a level of sensitivity indicates that Raman spectroscopy is proficient in decreasing the likelihood of missed diagnoses, enhancing its role as a reliable tool in the detection of laryngeal carcinoma. Furthermore, the specificity outcomes across these studies varied from 0.82 to 0.93, leading to a pooled specificity of 0.91, accompanied by a 95% confidence interval of 0.88–0.93(Fig. 4). This high level of specificity suggests that Raman spectroscopy is effective in minimizing false positive results, thereby increasing the accuracy of diagnoses in clinical settings. Together, these findings provide strong evidence of the utility of Raman spectroscopy as a valuable diagnostic method for laryngeal carcinoma, with significant implications for improving patient outcomes through more accurate detection and identification of the disease.

Fig. 4.

The forest plot of pooled sensitivity and pooled specificity

The combined positive likelihood ratio (LR) and negative LR were 9.70 (95% CI 7.13–13.21) and 0.13 (95% CI 0.11–0.16) (Fig. 5), respectively. The combined diagnostic odds ratio (DOR) was 4.13 (95% CI 3.90–4.72) (Fig. 6). The overall diagnostic performance was assessed using the summary receiver operating characteristics (SROC) curve, with the area under the curve (AUC) measuring 0.95 (Fig. 7). In our meta-analysis, the pre-test probability for laryngeal carcinoma patients was estimated at 11%, which corresponded to a positive predictive probability (PPP) of 91% (Fig. 8).

Fig. 5.

The forest plot of PLR and NLR

Fig. 6.

The forest plot of DOR

Fig. 7.

Summary receiver operating characteristic (SROC) plot of RS imaging was used to predict larynx cancer based on combination of sensitivity and specificity weighted for sample size of each data

Fig. 8.

Posttest probability of RS for detecting larynx cancer

Exploring heterogeneity

We conducted a meta-regression analysis to examine the potential heterogeneity among studies. The covariates considered included publication year (2000–2013 or 2014–2025), region (Asia or other), and diagnostic algorithm (PCA-LDA or others). The results indicated that the P-values for both region and diagnostic algorithm exceeded 0.05, as illustrated in Fig. 9, suggesting no significant heterogeneity between regions and diagnostic algorithms. This finding was further corroborated through subgroup analysis, as presented in Table 2. While no significant difference in sensitivity was observed over time, a statistically significant difference in specificity was identified, which was also supported by subgroup analysis, as shown in Table 2. This suggests that diagnostic accuracy may have improved over time.

Fig. 9.

Univariable meta-regression and subgroup analyses

Table 2.

Results of subgroup analysis

Factors	Pooled sensitivity (95%CI)	P value	Pooled specificity (95%CI)	P value
Overall studies	0.88 (0.86–0.90)		0.91 (0.88–0.93)
Year
2000–2013	0.95 (0.92–0.97)	0.99	0.90 (0.88–0.91)	0.02
2014–2025	0.95 (0.92–0.97)		0.95 (0.94–0.96)
Region
Asian	0.95 (0.92–0.97)	0.74	0.93 (0.91–0.95)	0.12
Other	0.94 (0.88–0.97)		0.94(0.92–0.96)
DA
PCA-LDA	0.96 (0.93–0.98)	0.16	0.89 (0.87–0.91)	0.08
Other	0.94 (0.92–0.96)		0.94 (0.93–0.96)

CI confidence interval, DA diagnostic algorithms, PCA-LDA: principle component analysis integrated with linear discriminant analysis

Discussion

Laryngeal carcinoma continues to pose a significant threat to individual’s quality of life, primarily due to insufficient early screening rates and limitations in diagnostic techniques. Traditional diagnostic methods often rely on invasive procedures such as endoscopic biopsies, which not only demand a high level of technical skill from the operator but also face challenges in detecting early microlesions [13]. Additionally, pathological diagnoses frequently carry a risk of false negatives, stemming from sampling errors or low patient cooperation. To address these challenges, researchers have developed novel clinical testing tools and biomarkers. Among these, Raman spectroscopy technology has gained prominence due to its advantages of non-invasive sampling, real-time analysis, and the capability to detect molecular vibration fingerprints. This technology identifies early lesions by capturing characteristic spectral changes in cellular metabolic products, demonstrating significant clinical potential in minimally invasive contexts [8, 14]. Consequently, it opens new avenues for the precise diagnosis and treatment of laryngeal carcinoma. This study aims to systematically evaluate the clinical diagnostic efficacy of Raman spectroscopy technology for laryngeal carcinoma, rigorously verifying its diagnostic accuracy through statistical methods. In terms of methodology, to ensure the robustness of the meta-analysis results, the research team employed a random effects model for data with low heterogeneity and a fixed effects model for studies exhibiting potential heterogeneity [10, 15]. This approach effectively mitigates the impact of diagnostic threshold effects and inter-study heterogeneity on the results. The primary diagnostic performance indicators reveal that the combined sensitivity reaches 88% (95% CI 0.86–0.90), indicating the technology’s ability to accurately identify 88% of laryngeal carcinoma patients. Additionally, the combined specificity is 91% (95% CI 0.88–0.93), meaning that 91% of healthy individuals can be correctly excluded. The analysis of the receiver operating characteristic curve reveals an area under the curve of 0.95, with a 95% confidence interval ranging from 0.94 to 0.98. This high AUC value suggests a significant level of reliability for the diagnostic test, as an AUC that approaches 1 indicates excellent diagnostic performance. Additionally, the combined DOR is calculated to be 4.13, accompanied by a 95% confidence interval of 3.90–4.72. This DOR not only encapsulates the effectiveness of the test in terms of both sensitivity and specificity but also implies that a higher DOR value corresponds to superior diagnostic discrimination, affirming the test’s capability to accurately differentiate between the presence and absence of the condition in question.

The combined sensitivity of 88% (95% CI 0.86–0.90) and specificity of 91% (95% CI 0.88–0.93), along with an area under the SROC curve of 0.95 (95% CI 0.94–0.98), indicates a robust diagnostic performance that exceeds the sensitivity threshold of traditional endoscopic biopsy, which typically is ≤75%. All seven utilized employed spontaneous Raman spectroscopy, though variations existed in laser wavelengths, tissue types (ex vivo/in vivo), and pretreatment procedures. Most utilized near-infrared lasers (785nm or 830 nm), with Lin2012 and Lin2016 being the only studies alongside in vivo measurements alongside ex vivo tissue analysis. To improve spectral quality, all research universally implemented smoothing filters and fluorescence background removal steps as standard pretreatment protocols.

It is recommended to integrate RS technology into the tertiary prevention system for laryngeal carcinoma. This integration should include the use of RS as an opportunistic screening tool in primary healthcare institutions for non-invasive screening of high-risk populations, such as those with a smoking index greater than 400, HPV positivity, or a history of occupational exposure [8, 16]. Additionally, it is advisable to establish a dual-track diagnostic pathway of ‘RS initial screening + pathological confirmation’ in secondary hospitals to guide stratified referrals based on positive likelihood ratios [11, 17]. Furthermore, in the realm of digital innovation in surgical paradigms, intraoperative RS real-time imaging systems can achieve molecular-level definitions of tumor boundaries, with a spatial resolution of 5 µm, thereby reducing diagnostic time compared to traditional frozen section diagnosis by 75%. It is suggested to develop an intelligent surgical navigation system guided by RS, which involves constructing a three-dimensional tumor metabolic model through multimodal RS images prior to surgery. Intraoperatively, dynamic margin assessment demonstrated a predictive value of RS for negative margins reaching 98% [18]. Postoperatively, the RS spectral database can be utilized for intelligent early warning of residual lesions. Notably, in laryngeal function-preserving surgery, RS technology quantitatively evaluates the biomechanical properties of the vocal membrane, significantly enhancing postoperative voice quality. The platformization of translational medicine research has allowed RS technology to evolve beyond a mere diagnostic tool, becoming an innovative engine for tumor biology research. It is recommended to establish an RS-driven laryngeal carcinoma research platform, develop machine learning-based metabolic pathway analysis algorithms, and achieve a closed-loop study from molecular phenotypes to clinical translation [9, 19]. A limitation of this study is the inclusion of literature solely in English, which may introduce language bias. Additionally, all included literature pertains to tissue specimens, with no relevant research on blood or other specimen types. Only seven high-quality studies were incorporated, indicating a need for more relevant literature to enhance the accuracy of this article. This study was conducted across subgroups. In the future, It is called to establish a standardized operating procedure (SOP) for Raman spectroscopy diagnosis of laryngeal carcinoma, including quality control indicators such as spatial resolution and laser power. More multi-center and large-sample studies are needed to verify the universality of the conclusions. Future studies will be expanded to non-English databases (such as CNKI, Wanfang, etc.) and non-invasive samples such as blood and saliva to enhance the clinical translational value.

Conclusion

In conclusion, RS is a tool with significant application prospects and value, which can be utilized for the screening and diagnosis of laryngeal carcinoma in the future. It will assist otolaryngologists in accurately locating tumors and making informed decisions, thereby effectively reducing the incidence laryngeal carcinoma and improving survival rates.

Author contributions

All authors contributed to the study conception and design. Data collection and preparation were performed by YH Z. Statistical analyses were performed by YH Z. The first draft of the manuscript was written by ZF Q. Writing reviewing and editing was performed by JZ G and W C. All authors read and approved the final manuscript.

Funding

This study was supported by grants from the Hospital-level Research Fund of Heping Hospital affiliated to Changzhi Medical College (HPJY202043).

Data availability

The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethical approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.