The accuracy of Fiber-Optic Raman Spectroscopy in the detection and diagnosis of head and neck neoplasm in vivo: a systematic review and meta-analysis

Search strategy

All studies were identified by systematically searching OVID EMBASE, OVID MEDLINE, Cochrane Library, and CNKI databases (up to February 2023), and there was no limit to the start date of the search. In this study Wen Chen, Yafei Chen and Chenzhou Wu performed the search strategy.

The authors display the details of the search strategy in Table 1.

Selection criteria

Studies were evaluated on the basis of the following criteria for inclusion: (1) only in vivo human samples of head and neck carcinomas were detected and diagnosed by fiber-optic RS. (2) All samples with head and neck carcinomas were investigated with histopathological diagnosis as the gold standard. (3) A healthy control group without head and neck carcinomas was included in the studies. (4) Data in the article can be used to construct a fourfold table including true positives (TPs), true negatives (TNs), false positives (FPs) and false negatives (FNs).

The exclusion criteria were as follows: (1) ex vivo sample detected, (2) studies that did not have a control group, and (3) reviews or duplicate reports.

Data extraction

We downloaded the full texts of all potential studies to ensure that they were eligible for inclusion. Three reviewers (Wen Chen, Yafei Chen and Chenzhou Wu) independently screened the 324 articles (title/abstract and full text). The whole screening process is blinded and the text software is used. Two reviewers independently extracted the data of each article and evaluated the quality of the article utilizing a standardized data extraction form. Disagreements were resolved by consensus. Data were collected as previously described in  Zhan et al. (2020), specifically the first author’s name, geographical location, demographic data (participants’ age and sex), tumor position, sample type, diagnostic algorithm, spectroscopy range, acquisition time, TP, TN, FP and FN.

Statistical analysis

All meta-analyses were performed in Stata 15.1 (Stata Corp, College Station, TX, USA).

The sensitivity, specificity, diagnostic threshold, diagnostic odds ratio (DOR), and 95% confidence interval (CI) were calculated to obtain the diagnostic accuracy of fiber-optic RSfor head and neck carcinomas. Outcome data were subject to statistical pooling through random effect models, which suggests that the studies from populations may affect the final results (Melsen et al., 2014; Lean et al., 2009). Also, we used the midas module to calculate summary statistics and SROC. The commands were “midas TP, TN, FP, FN, res(all)” and “midas TP, TN, FP, FN, plot sroc(both)”, respectively.

A summary receiver operating characteristic (SROC) curve and threshold analysis were carried out to investigate the threshold. The area under the curve (AUC) was calculated to evaluate the overall effectiveness of fiber-optic RS. If the SROC curves exhibited a shoulder peak, it indicated that thresholds may have an impact on the result. The diagnostic effect was excellent when the AUC value was between 0.9 and 1, favorable when the AUC value was between 0.8 and 0.9, fair when the AUC value was between 0.7 and 0.8, and poor when the AUC value was between 0.6 and 0.7. The diagnostic method was considered to have failed when the AUC fell between 0.5 and 0.6 (Metz, 1978).

The Q statistic and the inconsistency index (I²) statistic were used to further investigate heterogeneity. The Q statistic was used to illustrate the presence or absence of heterogeneity, and the I² index was used to classify the degree of heterogeneity (Huedo-Medina et al., 2006). The degree of heterogeneity was considered to be significant when the I² index was greater than 50% and the P value was less than 0.05  (Higgins et al., 2003). Subgroup analyses were performed for substantial heterogeneity. Publication bias was estimated by Deeks’ funnel plot asymmetry test, which was considered to exist when the P value was less than 0.05 (Begg & Mazumdar, 1994).

Quality assessment

The Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2) guidelines were used to systematically assess the quality of the studies included in this meta-analysis (high, unclear, or low) (Whiting et al., 2011). The main items included (1) patient selection, (2) the index test, (3) the reference standard and (4) flow and timing. The risk of bias was rated as low risk, high risk or unclear risk. The QUADAS-2 was performed by Review Manager 5.4. The quality of the included studies was evaluated independently by two reviewers (Yafei Chen and Chenzhou Wu) according to the QUADAS-2 guidelines. Disagreements were resolved by a third reviewer (Wen Chen).

Study selection and description of studies included in the article

Initially, the authors searched 658 articles from OVID EMBASE, OVID MEDLINE, the Cochrane Library and CNKI databases. After removing duplicates, 324 articles were selected. Then, 86 articles were identified after screening the titles and abstracts. Finally, 10 eligible articles were included in this meta-analysis. The full study screening and selection process is presented in Fig. 1.

Sixteen groups of data from 10 articles were extracted from these articles due to their characteristics (such as different spectrum ranges and different acquisition times), and none of these 16 groups of data were duplicated. Of all eligible studies (Žuvela et al., 2019; Lin et al., 2016b; Lin et al., 2017; Malik et al., 2017; Krishna et al., 2014; Lin, Cheng & Huang, 2012; Singh et al., 2013; Sahu et al., 2015; Ming et al., 2017; Lin et al., 2018), all articles were written in English. Among the 16 groups of data, the sample types included the larynx (n = 4) (Lin et al., 2016b; Lin, Cheng & Huang, 2012), the nasopharynx (n = 8) (Žuvela et al., 2019; Lin et al., 2017; Ming et al., 2017; Lin et al., 2018) and the oral cavity (n = 4) (Malik et al., 2017; Krishna et al., 2014; Singh et al., 2013; Sahu et al., 2015). The Raman spectral range applied in all eligible studies was divided into three categories, including the fingerprint region (FP) (n = 8) (Žuvela et al., 2019; Lin et al., 2016b; Lin et al., 2017; Malik et al., 2017; Krishna et al., 2014; Singh et al., 2013; Sahu et al., 2015; Lin et al., 2018), high wavenumber (HW) ( n = 4) (Žuvela et al., 2019; Lin et al., 2016b; Lin et al., 2017; Lin, Cheng & Huang, 2012) and FP + HW (n = 4) (Žuvela et al., 2019; Lin et al., 2016b; Lin et al., 2017; Ming et al., 2017). Considering the variable Raman instrumentation used in the ten studies, acquisition times were divided into two groups: acquisition times ≤ 1 s (n = 12) (Žuvela et al., 2019; Lin et al., 2016b; Lin et al., 2017; Lin, Cheng & Huang, 2012; Ming et al., 2017; Lin et al., 2018) and acquisition times > 1 s (n = 4) (Malik et al., 2017; Krishna et al., 2014; Singh et al., 2013; Sahu et al., 2015). The details of each group of data are shown in Table 2. Partial least squares-discrimination analysis (PLS-DA), Leave-one-out cross-validation (LOOCV), principal component analysis + Linear discriminant analysis (PCA + LDA), genetic algorithm-partial least squares-linear discriminant analysis (GA-PLS-LDA), and stepwise analysis of multiple linear regression (SMLR) in Table 2 refer to different diagnostic algorithms of Raman spectra.

Pooled results

Overall results

Ten studies (Žuvela et al., 2019; Lin et al., 2016b; Lin et al., 2017; Malik et al., 2017; Krishna et al., 2014; Lin, Cheng & Huang, 2012; Singh et al., 2013; Sahu et al., 2015; Ming et al., 2017; Lin et al., 2018) (16 groups of data) were included. In these studies, a total of 5365 in vivo Raman spectra (cancer = 1,746; normal = 3,619) were acquired from 877 patients. Their coalescent sensitivity and specificity results for fiber-optic RS were 0.88 (95% CI [0.84–0.91], P = 0.00, I² = 85.88) and 0.94 (95% CI [0.91–0.96], P = 0.00, I² = 76.64), respectively. The DOR was 105.69 (95% CI [67.50–165.47], P = 0.00, I² = 100.00). SROC curves were generated to estimate the overall diagnostic accuracy, and the AUC was 0.96 (95% CI [0.94–0.97]).

Site of disease

Larynx cancer

Two of the included studies (Lin et al., 2016b; Lin, Cheng & Huang, 2012) (4 groups of data) assessed larynx samples. A total of 397 in vivo Raman spectra (cancer = 162; normal = 235) were acquired from 99 patients. The coalescent sensitivity and specificity results for fiber-optic RS were 0.88 (95% CI [0.81–0.92], P = 0.19, I² = 36.52) and 0.88 (95% CI [0.83–0.92], P = 0.85, I² = 0.00), respectively. The DOR was 47.63 (95% CI [21.93–103.45], P = 0.275, I² = 22.6). SROC curves were generated to estimate the overall diagnostic accuracy, and the AUC was 0.92(95% CI [0.89–0.94]).

Nasopharyngeal cancer

Four of the included studies (Žuvela et al., 2019; Lin et al., 2017; Ming et al., 2017; Lin et al., 2018) (8 groups of data) assessed nasopharynx samples. A total of 1298 in vivo Raman spectra (cancer = 580; normal = 718) were acquired from 296 patients. Their coalescent sensitivity and specificity results for fiber-optic RSwere 0.88 (95% CI [0.83–0.91], P = 0.02, I² = 57.83) and 0.97 (95% CI [0.90–0.99], P = 0.00, I² = 72.94), respectively. The DOR was 118.07 (95% CI [71.38–195.30], P = 0.294, I² = 17.3). SROC curves were generated to estimate the overall diagnostic accuracy. The AUC was 0.94 (95% CI [0.92–0.96]).

Oral cancer

Four of the included studies (Malik et al., 2017; Krishna et al., 2014; Singh et al., 2013; Sahu et al., 2015) (4 groups of data) assessed oral samples. A total of 3670 in vivo Raman spectra (cancer = 1,004; normal = 2,666) were acquired from 482 patients. Their coalescent sensitivity and specificity results for fiber-optic RS were 0.87 (95% CI [0.76–0.94], P = 0.00, I² = 95.94) and 0.93 (95% CI [0.90–0.95], P = 0.00, I² = 88.50), respectively. The DOR was 90.13 (95% CI [32.91–246.86], P = 0.000, I² = 93.6). SROC curves were generated to estimate the overall diagnostic accuracy, and the AUC was 0.96 (95% CI [0.94–0.97]).

Raman spectral range

Fingerprint range

Eight of the included studies (8 groups of data) (Žuvela et al., 2019; Lin et al., 2016b; Lin et al., 2017; Malik et al., 2017; Krishna et al., 2014; Singh et al., 2013; Sahu et al., 2015; Lin et al., 2018) assessed FP samples. A total of 4371 in vivo Raman spectra (cancer = 1,314; normal = 3,057) were acquired from 325 patients. Their coalescent sensitivity and specificity results for fiber-optic RS were 0.87 (95% CI [0.80–0.91], P = 0.00, I² = 91.70) and 0.93 (95% CI [0.90–0.95], P = 0.00, I² = 81.45), respectively. The DOR was 85.36 (95% CI [43.75–166.55], P = 0.000, I² = 86.1). SROC curves were generated to estimate the overall diagnostic accuracy. The AUC was 0.96 (95% CI [0.94–0.97]).

High wavenumber range

Four of the included studies (4 groups of data) (Žuvela et al., 2019; Lin et al., 2016b; Lin et al., 2017; Lin, Cheng & Huang, 2012) assessed HW samples. A total of 512 in vivo Raman spectra (cancer = 241; normal = 271) were acquired from 256 patients. The coalescent sensitivity and specificity results for fiber-optic RSwere 0.86 (95% CI [0.81–0.91], P = 0.17, I² = 41.00) and 0.94 (95% CI [0.82–0.98], P = 0.01, I² = 71.81), respectively. The DOR was 66.59 (95% CI [24.19–183.29], P = 0.102, I² = 51.7). SROC curves were generated to estimate the overall diagnostic accuracy, and the AUC was 0.91 (95% CI [0.88–0.93]).

Fingerprint range + high wavenumber range

Four of the included studies (4 groups of data) (Lin et al., 2016b; Lin et al., 2017; Ming et al., 2017; Lin et al., 2018) assessed FP + HW samples. A total of 482 in vivo Raman spectra (cancer = 191; normal = 291) were acquired from 296 patients. The coalescent sensitivity and specificity results for fiber-optic RSwere 0.93 (95% CI [0.88–0.96], P = 0.97, I² = 0.00) and 0.96 (95% CI [0.88–0.98], P = 0.04, I² = 63.60), respectively. The DOR was 199.73 (95% CI [89.96–443.45], P = 0.493, I² = 0.0). SROC curves were generated to estimate the overall diagnostic accuracy, and the AUC was 0.94 (95% CI [0.91–0.96]).

Acquisition time

≤1 s

Six of the included studies (Žuvela et al., 2019; Lin et al., 2016b; Lin et al., 2017; Lin, Cheng & Huang, 2012; Ming et al., 2017; Lin et al., 2018) (12 groups of data) assessed acquisition times ≤1s. A total of 1,695 in vivo Raman spectra (cancer = 742; normal = 953) were acquired from 395 patients. Their coalescent sensitivity and specificity results for fiber-optic RSwere 0.88 (95% CI [0.85–0.91], P = 0.03, I² = 49.80) and 0.95 (95% CI [0.90–0.97], P = 0.00, I² = 67.04), respectively. The DOR was 92.21 (95% CI [56.49–150.52], P = 0.101, I² = 36.2). SROC curves were generated to estimate the overall diagnostic accuracy, and the AUC was 0.94 (95% CI [0.91–0.95]).

>1 s

Four of the included studies (Malik et al., 2017; Krishna et al., 2014; Singh et al., 2013; Sahu et al., 2015) (4 groups of data) assessed acquisition times > 1s. A total of 3670 in vivo Raman spectra (cancer = 1,004; normal = 2,666) were acquired from 482 patients. Their coalescent sensitivity and specificity results for fiber-opticRS were 0.87 (95% CI [0.76–0.94], P = 0.00, I² = 95.94) and 0.93 (95% CI [0.90–0.95], P = 0.00, I² = 88.50), respectively. The DOR was 90.13 (95% CI [32.91–246.86], P = 0.000, I² = 93.6). SROC curves were generated to estimate the overall diagnostic accuracy, and the AUC was 0.96 (95% CI [0.94–0.97]).

Assessment of study quality

All QUADAS-2 items were used to estimate the eligible studies. The risk of bias of the eligible studies is presented in Fig. 2. We can see that all studies were judged as “high risk” on flow and timing domain relating to bias, which is irrational. The reason is that in these 10 studies, all “healthy tissue” has not been performed with pathological examination, while all “cancer tissue” have performed with pathological examination, for the ethical reasons. So, the answer of all the studies is “no” on the signaling question “did all patients receive the same reference standard”. Regardless of this issue, most risk assessments were considered “low risk”.

Publication bias and heterogeneity

The forest plot of the sensitivity and specificity of each eligible study is shown in Fig. 3, and indicates that the heterogeneity was significant. In addition, the Q test values of the sensitivity and specificity were 106.23 (P = 0.00) and 64.21 (P = 0.00), respectively, and the I2 index of the sensitivity and specificity were 85.88 (95% CI [79.99–91.77]) and 76.64 (95% CI [65.45–87.83]), respectively. The results of heterogeneity in each subgroup are presented in Table 3.

No significant publication bias was found in this meta-analysis by Deeks’ funnel plot asymmetry test. The funnel plot is shown in Fig. 4.