Training and validation of a 1D CNN model for accurate three-class classification of oral cancer using serum Raman spectroscopy

Abstract

In this study, a one-dimensional convolutional neural network (1D-CNN) was developed for classifying serum Raman spectra into three oral cancer-related classes: normal, precancerous, and cancerous. To address class imbalance, class-weighted learning was employed during training, while augmentation was evaluated separately as a comparative strategy. The final selected non-augmented model achieved an overall testing accuracy of 83%, with a macro-precision of 85%, macro-recall of 84%, and macro-F1 score of 83%, indicating balanced three-class performance. To further assess real-world applicability, an independent test dataset collected later from the same hospital was used. This cohort followed the same acquisition and preprocessing protocol but consisted of entirely new patients. The model maintained consistent classification performance on this independent dataset, supporting its generalizability within the same clinical environment. These findings highlight the potential of combining Raman spectroscopy with deep learning as a non-invasive framework for oral cancer-related biosample classification.

Introduction

Oral cancer remains one of the leading causes of morbidity and mortality worldwide, with late-stage diagnosis being a critical challenge that significantly reduces survival rates [1]. Globally, in 2020 there were an estimated 377,713 new cases of oral cancer (13th most common cancer worldwide) and 177,757 deaths [2]. The most common form of oral cancer is oral squamous cell carcinoma (OSCC). Oral potentially malignant disorders (OPMDs), such as leukoplakia, oral lichen planus, and erythroplakia, are oral mucosal lesions with a higher risk of progressing to oral cancer. Early recognition of OPMDsis critical while plays a pivotal role in improving patient outcomes, but frontline healthcare providers—including primary care physicians, nurses, dentists, and hygienists—often lack the training to reliably distinguish OPMDs and OSCC from benign lesions [3]. Conventional diagnostic methods, including histopathology and biopsy, remain the gold standard but are invasive, time-consuming, and resource-intensive. These barriers, combined with gaps in knowledge and weak communication with specialists, contribute significantly to delayed diagnosis and treatment. This has prompted growing interest in the development of non-invasive, rapid, and accurate diagnostic methods for early-stage screening.

Raman spectroscopy has emerged as a promising label-free, optical technique capable of providing detailed biochemical fingerprints of biological samples, including serum, saliva, and tissue [4–8]. It has emerged as a promising tool in clinical research, particularly for the identification and diagnosis of oral cancer [9]. It offers high sensitivity to molecular changes associated with malignancy and has shown potential for cancer detection across various tissue types. However, the complex and high-dimensional nature of Raman spectral data poses challenges for traditional statistical analysis, particularly when dealing with small and imbalanced clinical datasets. In this study, serum was chosen as a reliable and stable biofluid for Raman spectroscopic analysis. Although both serum and saliva show a progressive increase in sialic acid levels from normal to OPMD and oral cancer [10], serum provides a more robust and consistent biomarker profile due to its higher sialic acid concentration and lower susceptibility to external influences such as oral hygiene, diet, circadian rhythm, and microbial activity [11]. Compared with other biofluids like saliva or urine, blood and serum-based assays offer greater stability, reproducibility, and ease of repeated sampling [12]. Moreover, serum better reflects systemic tumor-associated biochemical changes, making it particularly suitable for accurate and reproducible spectral classification [13].

With the increasing application of Raman spectroscopy, data-driven deep learning methods have gained wide attention. End-to-end approaches such as Convolutional Neural Networks (CNNs) have been successfully applied to spectral classification, with progress reported in different fields [14]. These advances highlight the strengths of deep learning in signal extraction, feature learning, and modeling complex spectral relationships. However, the integration of Raman spectroscopy with deep learning still faces challenges, particularly due to the ’black-box’ nature of these models, which limits interpretability and raises concerns for clinical deployment [15]. Reliable and explainable predictions are therefore essential to build trust in such systems. Deep learning methods such as CNNs have demonstrated superior performance in spectral analysis and biomedical signal classification [16]. Unlike classical machine learning methods, CNNs can directly learn hierarchical patterns from raw or preprocessed input, making them well-suited for biomedical signal analysis [17]. However, the limited availability of biomedical spectral datasets often leads to class imbalance and overfitting. Data augmentation techniques, including jittering, scaling, and shifting, have been widely employed to expand training sets and improve generalization [18–20]. Although generative AI techniques are advancing rapidly, data augmentation remains a simpler and well-established strategy in spectroscopy, offering an effective means to enhance limited datasets and improve model generalization [21].

In this study, we present a 1D-CNN architecture designed to classify serum Raman spectra into three diagnostic categories relevant to oral cancer: normal, precancer, and cancer. To address class imbalance, a class-weighted learning strategy was adopted during CNN training, while spectral augmentation was evaluated separately as a comparative strategy. To assess the effect of augmentation, the classification performance of the 1D-CNN model was compared under two settings: training without augmentation and training with augmentation. Under the present experimental configuration, the non-augmented model achieved better overall performance. In addition to internal testing through repeated random hold-out evaluation, we employed a temporally separated independent test set comprising newly collected samples from the same hospital to assess real-world generalization. This design ensured that the model encountered previously unseen cases reflective of a practical clinical environment. By integrating Raman spectroscopy with deep learning, this study aims to advance non-invasive approaches for early detection and risk stratification of oral cancer.

Materials and methods

Sample preparation and data acquisition

This study was approved by the Institutional Review Board (Chang Gung Medical Foundation, Taiwan; IRB Nos. 202300682B0 and 201800439B0). Written informed consent was obtained from all participants. The study was carried out at the Department of Otolaryngology Head and Neck Surgery, Chang Gung Memorial Hospital, and enrolled patients with histologically confirmed oral lesions, including 42 (oral squamous cell carcinoma, OSCC) and 39 precancerious cases. Serum samples were collected from participants representing three diagnostic categories: class1 (Cancer, n = 42), class 2 (Precancer, n = 39), and class 3 (Normal = 20). The demographic details of the study cohorts are presented in Table 1. The analyzed dataset comprised 210 spectra from cancer patients, 195 spectra from precancer cases, and 100 spectra from healthy subjects. After averaging the multiple measurements obtained for each subject, the final dataset consisted of 42 cancer, 39 precancer, and 20 healthy samples. All blood serum samples were collected in sterile plastic serum collection tubes without anticoagulant. After clot formation, samples were centrifuged at 3000 rpm for 15 min to separate the serum. The isolated serum was aliquoted into sterile microcentrifuge tubes and stored at - 80 C for long-term stability until further Raman analysis. Before analysis, samples were thawed under controlled conditions and processed according to the experimental protocol. Serum Raman spectra were recorded using a confocal Micro Raman Identify system (ProTrusTech Co., Ltd., Tainan, Taiwan) equipped with a 532 nm excitation laser. The nominal maximum laser power of the instrument is approximately 126 mW; however, in the present study, the laser power at the sample was limited to 31.5 mW to avoid photothermal effects and sample degradation. Before each measurement session, wavelength calibration was performed using a silicon wafer by confirming the characteristic Raman peak at 520 . Spectra were collected with an integration time of 10 s and 3 s accumulations for each sample. For each patient, five Raman measurements were acquired and averaged to obtain one representative spectrum per subject. This strategy, consistent with prior spectroscopic studies, ensured that classification was performed at the patient level, thereby reducing intra-sample variability and avoiding bias from repeated measurements [22].

Table 1.

Clinical and demographic profile of the study cohort ()

Variable	Category	Age (mean ± SD)	Gender (M:F)
Overall	All Samples		93 : 8
Group	Malignant		37 : 5
	Premalignant		36 : 3
	Normal		20 : 0
Statistics	p-value	0.027	–
Total Count		101

Data preprocessing and analysis

All spectral preprocessing and data analysis were carried out in MATLAB (R2018a, MathWorks, USA). To reduce high-frequency noise and improve spectral smoothness, the raw Raman spectra were first filtered using a third-order Savitzky–Golay smoothing algorithm. Baseline correction was then applied to remove background contributions and fluorescence-related drift. Following this, the spectra were normalized using an area-under-the-curve (AUC) approach to minimize intensity variation and reduce data redundancy across samples. In this normalization procedure, each spectral intensity value was scaled relative to the total integrated area of the corresponding spectrum, thereby standardizing the spectra prior to subsequent analysis. After preprocessing, principal component analysis (PCA) was applied for dimensionality reduction. PCA was fitted using the training data only, and the number of retained principal components was selected to preserve more than 99% of the cumulative variance. The testing data were then projected into the same PCA space using the coefficients derived from the training set. In the present study, we analyzed serum Raman spectra from 42 oral cancer, 39 precancer, and 20 healthy cases, representing an inherently imbalanced dataset. Addressing class imbalance is important for reliable classification performance. In general, imbalance may be handled using data-level approaches, such as undersampling, oversampling, or augmentation, or algorithm-level approaches, such as class weighting or threshold adjustment [5]. In this study, class imbalance was primarily addressed through a class-weighted learning strategy while retaining all original samples. The dataset was partitioned at the patient level into training (70%) and testing (30%) subsets, and the 1D-CNN model was trained on the original training data and evaluated on the corresponding unseen test set. This procedure was repeated across multiple random hold-out runs to improve robustness and reduce dependence on a single data split. Data augmentation was additionally evaluated as a comparative strategy, but the non-augmented configuration yielded better overall performance. Finally, an out-of-distribution cohort collected later from newly recruited patients was used for independent evaluation of model generalizability.

Data augmentation

To expand the training dataset and mitigate overfitting, each training sample was augmented using jittering (Gaussian noise addition), scaling (multiplicative intensity variation), and shifting (spectral axis translation). These methods were chosen to simulate realistic Raman acquisition variability while preserving the underlying biochemical structure of the spectra. In particular, shifting was used to model small Raman shift variations during acquisition and to improve the robustness of CNN-based feature extraction to minor peak-position changes. The dataset was first partitioned into training and testing subsets at the patient level to ensure complete separation between model development and evaluation. Data augmentation was then applied exclusively to the training set, where each original spectrum was augmented eight times, resulting in a ninefold expansion of the training data, while the testing set remained unchanged. The augmentation was implemented via a custom function (augmentSpectrum) that applied Gaussian noise with a noise level of 0.01, multiplicative scaling in the range 0.95-−1.05, and spectral shifting in the range of to Raman units. These parameters were empirically selected to introduce realistic spectral variability while preserving the underlying biochemical characteristics of the original spectra. However, an additional comparative analysis showed that, under the present experimental setting, the model trained on the original non-augmented spectra achieved better overall performance than the augmented-data model. This suggests that the selected augmentation strategy did not improve generalization for the current dataset and may have introduced synthetic variability that was less representative of the true Raman spectral distribution.

CNN architecture design

A one-dimensional Convolutional Neural Network (1D-CNN) was developed for classification of Raman spectra into three diagnostic categories as shown in Fig. 1. The PCA-transformed spectra were used as input to the 1D-CNN model. The network architecture comprised three sequential convolutional blocks, each including a convolution layer with 64 filters and a kernel size of , followed by batch normalization and Leaky ReLU activation. Max-pooling layers with a pooling size of and a stride of 1X4 were applied after the first two convolutional blocks to progressively reduce spectral dimensionality while preserving relevant features.

Fig. 1.

Architecture of the one-dimensional convolutional neural network (1D-CNN) used for three-class classification of serum Raman spectra (Normal, Precancer, and Cancer)

Following convolutional feature extraction, the model included a fully connected layer with 64 neurons, followed by Leaky ReLU activation and a dropout layer with a rate of 0.5 to reduce overfitting. A final fully connected layer with three output neurons and softmax activation was used for three-class classification. The architecture of the proposed 1D-CNN model is summarized in Table 2. The model was trained using the Adam optimizer with an initial learning rate of 0.001, a mini-batch size of 4, and a maximum of 10 epochs. Overall, the architecture was designed to balance efficient feature extraction with robust generalization for a relatively small and imbalanced Raman dataset.

Table 2.

Architecture of the 1D-CNN model designed for three-class classification of serum Raman spectra. The network utilizes kernels to perform convolutions exclusively along the spectral axis, ensuring feature extraction is focused on vibrational modes

Model	Type	Filters	Size	Strides	Epoch
1D-CNN	Conv layer	64	[1 8]	1	10
	BatchNorm	–	–	–
	LeakyReLU	–	–	–
	MaxPooling2D	–	[1 2]	[1 4]
	Conv layer	64	[1 8]	1
	BatchNorm	–	–	–
	LeakyReLU	–	–	–
	MaxPooling2D	–	[1 2]	[1 4]
	Conv layer	64	[1 8]	1
	BatchNorm	–	–	–
	LeakyReLU	–	–	–
	Fully Connected	64	–	–
	LeakyReLU	–	–	–
	Dropout	–	–	–
	Fully Connected	3	–	–
	Softmax	–	–	–
	Classification	–	–	–

Model evaluation

For multi-class classification, precision, recall, and F1-score are considered important complementary metrics because they provide class-sensitive performance assessment beyond overall accuracy [23]. Precision measures the proportion of predicted samples in a class that are truly correct, whereas recall measures the proportion of actual samples in a class that are correctly identified. The F1-score provides a balanced measure by combining precision and recall. A confusion matrix was constructed for each run, from whcih accuracy, precision, recall, and F1-score were calculated for each class. For multi-class evaluation, macro-precision, macro-recall, and macro-F1 were also calculated by averaging performance equally across classes, while weighted-F1 was computed by accounting for class size. The corresponding calculation formulas are presented in Eqs. (1)–(8). where , , and denote the true positives, false positives, and false negatives for class , respectively; is the total number of classes; is the number of samples in class ; and is the total number of samples. Model performance was evaluted on the corresponding unseen test set using multiple classification metrics. Training and testing accuracy and loss were recorded across epochs, and the mean ± standard deviation (SD) curves over repeated runs were used to assess convergence behavior, robustness, and model generalization.

1

2

3

4

5

6

7

8

Independent validation data set

For independent validation, we employed a temporally separated dataset collected at a later period from the same hospital. The acquisition protocol, instrumentation, and preprocessing steps were consistent with those used for the training and testing datasets. However, all validation samples were obtained from newly recruited patients, ensuring no overlap with the original dataset and reflecting a real-world scenario where the model encounters unseen cases in the same clinical environment.

Results and discussion

Spectral assignments

The recorded Raman spectra from serum samples exhibited several characteristic vibrational modes associated with biomolecular changes in oral cancer progression, can be seen in Fig. 2. Prominent peaks were observed in the regions corresponding to proteins (amide I and III bands), lipids, and nucleic acids. Comparative inspection revealed subtle but consistent variations between the three diagnostic classes. In the Normal group, characteristic carotenoid peaks at 1006 , 1156 and 1524 , along with a collagen-associated band at 1205 , were more prominent compared to the precancer group. Comparative analysis of the average Raman spectra for normal and cancer groups revealed significant variations in both peak intensities and frequency shifts, particularly in regions associated with protein amide bands and lipid vibrations. While normal samples exhibited stronger carotenoid and collagen signals, the cancer group demonstrated enhanced features at 754 , 1450 and 1656 , corresponding to heme, CH2 bending vibrations, and the amide I band, respectively. Similarly, the cancer versus precancer spectra displayed distinct spectral features dominated by protein and phospholipid bands, again with notable peaks at 754 , 1450 , and 1656 . Additional tryptophan-related bands appeared between 1306–1346 , characterized by overlapping contributions from multiple biomolecules [24]. These included CH bending and twisting vibrations in DNA/RNA, lipid acyl chains, protein side chains, and nucleic acid bases such as guanine. Key spectral hallmarks of cancer samples were a broad and intense amide I band at 1656 and a strong CH2 bending vibration at 1450 , reflecting alterations in protein secondary structure and lipid content [25]. The Raman peak observed near 1244 corresponds to the Amide III band, which arises from vibrations of the protein backbone [26]. This mode primarily involves a combination of C–N stretching and N–H bending. The precise position and intensity of the Amide III band are highly sensitive to the secondary structure of proteins, reflecting contributions from alpha-helices, beta-sheets, and random coil conformations. In serum samples, this band serves as a strong indicator of overall protein content and structure, particularly reflecting the presence of major proteins such as albumins and globulins. All assignments of peaks are summarized in Table 3.

Fig. 2.

Average Raman spectra of serum collected from normal (blue curve), precancer (red curve) and cancer patients suffering from oral cancer (black curve)

Table 3.

Key Raman spectral peaks and their biochemical assignments in serum samples

Peak position (cm)	Band assignment	Vibrational mode	References
754	Tryptophan	Ring breathing, protein marker	[27]
1006	Carotenoids	Ring breathing	[28]
1156	Carotenoids	C–C stretching	[28]
1205	Collagen	C–CH vibration	[29]
1244	Amide III	C–N stretching, N–H bending	[26]
1306–1346	Tryptophan / nucleic acids	CH bending, twisting	[24]
1450	CH bending	C–H deformation	[25]
1524	Carotenoids	C=C stretching	[30]
1656	Amide I	C=O stretching, C–N vibration	[31]

In Fig. 3, it is evident that several Raman bands exhibit intensity differences across the cancer, precancer, and normal serum groups. Notably, the carotenoid associated peaks at 1006 , 1156 , and 1524 show higher intensities in the normal group compared to precancer and cancer patients, reflecting a depletion of carotenoids during disease progression. Similarly, the collagen-associated band at 1244 (Amide III) demonstrates reduced intensity in cancer and precancer sera relative to normal controls, suggesting structural alterations in serum proteins. Conversely, bands at 1450 (CHbending vibration) and 1655 (Amide I), which are primarily linked to lipids and proteins, display relatively stronger signals in cancer patients, indicating biochemical changes in lipid and protein composition during cancer transformation. Smaller differences can also be observed at 753 , 963 , 1309 , and 1343 , further reflecting disease-associated perturbations in nucleic acids, proteins, and lipids. Overall, the spectral variations across these diagnostic bands highlight biochemical differences between cancer, precancer, and normal groups, particularly involving carotenoids, lipids, and proteins, which can serve as potential biomarkers for oral cancer detection.

Fig. 3.

Average ± standard deviation (SD) value of Raman intensity of bands observed in different class of serum samples

Classification performance

The training and testing curves indicate stable learning behavior of the proposed 1D-CNN model over 10 epochs. The model was trained for 10 epochs because both the accuracy and loss curves showed rapid convergence during the early training stage, while performance stabilized in the later epochs. Extending training beyond this point was unlikely to provide substantial improvement and could increase the risk of overfitting due to the limited dataset size.

For the augmented model, the mean training accuracy increased from approximately 40% to 85–90%, while the mean testing accuracy improved from around 30% to 75–76% [Fig. 4(a)]. Similarly, the mean training and testing loss values decreased from about 2.0 to below 0.5 and 0.7, respectively [Fig. 4(b)]. These trends indicate progressive convergence without severe divergence between training and testing performance. For the non-augmented model, both training and testing accuracy also increased steadily, while the corresponding loss values decreased across epochs [Figs. 4 (c) and 4(d)]. By the final epochs, the testing accuracy reached approximately 82–85%, and the testing loss decreased to around 0.60, suggesting stable learning and good generalization. Compared with the augmented model, the non-augmented model achieved better final testing performance, indicating that augmentation improved training smoothness but did not provide additional generalization benefit under the present experimental setting.

Fig. 4.

Performance metrics of the proposed 1D-CNN model: a accuracy and b loss with data augmentation; c accuracy and d loss without data augmentation. Results represent the mean and SD (error bars) over five repeated runs

Tables 4 and 5 summarize the classification performance of the augmented and non-augmented models, respectively. In the non-augmented model, the normal class showed the strongest overall performance, with precision, recall, and F1-score values of , , and , respectively. The cancer class also demonstrated strong discrimination, with precision, recall, and F1-score values of , , and . In contrast, the precancer class remained the most challenging category, showing comparatively lower recall and F1-score , likely due to partial spectral overlap with both normal and cancer groups. Specifically, the final selected non-augmented model improved the augmented model by 7.9%, 8.2%, and 7.3% in terms of macro-F1, macro-precision, and macro-recall, respectively, while the overall accuracy increased by 6.3%. Overall, these updated results confirm that the non-augmented model outperformed the augmented configuration and was therefore selected as the final model for the independent validation of the model.    

Table 4.

Comprehensive classification performance metrics of the proposed 1D-CNN model trained with data augmentation, including per-class results and global averaged measures

Metric / Class	Precision	Recall	F1-score
Cancer
Precancer
Normal

Global Metrics	Value	Metric	Value
Accuracy		Macro-Precision
Macro-Recall		Macro-F1
Weighted-F1

Table 5.

Comprehensive classification performance metrics of the proposed 1D-CNN model trained without data augmentation, including per-class results and global averaged measures

Metric /Class	Precision	Recall	F1-score
Cancer
Precancer
Normal

Global Metrics	Value	Metric	Value
Accuracy		Macro-Precision
Macro-Recall		Macro-F1
Weighted-F1

Independent validation results

To further validate the robustness of the model, an independent test cohort was subsequently obtained from Chang Gung Memorial Hospital, Taiwan. This specific dataset was collected and analyzed at a later stage, separate from the initial training and testing phases, to ensure the model’s predictive performance remains consistent across different temporal sampling points. The independent test cohort consisted of 6 newly recruited subjects, including 2 cancer, 2 precancer, and 2 normal cases. Detailed demographic information, including age and sex, is provided in Table 6. The inclusion of demographic information improves the clinical interpretability of the independent evaluation cohort. The independent test results reported here correspond to the final selected 1D-CNN model trained without data augmentation, since this configuration yielded better overall classification performance than the augmented model. Overall, the model assigned relatively high probabilities to the correct classes for most samples. For example, cancer cases were correctly classified with cancer probabilities of 0.88 and 0.93, indicating strong model confidence. Normal samples were also identified reliably, with predicted probabilities exceeding 0.78 for the normal class. In comparison, precancerous samples showed lower classification confidence. One precancer case was misclassified as cancer with a probability of 0.66, whereas another was correctly classified as precancer with a probability of 0.73. These findings suggest that the model maintained reasonable generalization on unseen patient samples, while the main source of ambiguity remained the boundary between precancer and cancer. This trend is consistent with the partial spectral overlap observed during internal testing. However, these findings should be interpreted as preliminary because of the limited size of the independent cohort.

Table 6.

The independent test results reported here correspond to the final selected 1D-CNN model trained without data augmentation, since this configuration yielded better overall classification performance than the augmented model

Sample ID	True Class	Age(Years), Gender	Probability	Probability	Probability	Correct
			(Cancer)	(Precancer)	(Normal)	correctly classified
P1	Cancer	55,M	0.8827	0.1067	0.0106	Yes
P2	Precancer	60,M	0.6554	0.2125	0.1321	No
P3	Normal	58,M	0.0923	0.0545	0.8532	Yes
P4	Cancer	59,M	0.9335	0.0258	0.0407	Yes
P5	Precancer	54,M	0.2532	0.7310	0.0158	Yes
P6	Normal	70,F	0.0092	0.2074	0.7834	Yes

The proposed 1D-CNN effectively distinguished normal, precancer, and cancer serum spectra, demonstrating stable multi-class classification performance. Across repeated runs, the final selected non-augmented model achieved an overall accuracy of , with macro-precision, macro-recall, and macro-F1 values of , , and , respectively. These results indicate balanced discrimination across the three diagnostic groups, although the precancer class remained relatively more challenging, likely due to partial spectral overlap with both normal and cancer spectra. Overall, the findings confirm that the proposed deep learning framework was able to capture subtle biochemical variations in serum Raman spectra associated with progressive malignant transformation. Data augmentation using jittering, scaling, and spectral shifting was additionally evaluated as a comparative strategy because such approaches are widely used in spectroscopy to increase training diversity and reduce overfitting under limited-data conditions [32, 33]. However, under the present experimental setting, the non-augmented model outperformed the augmented model, suggesting that the synthetic perturbations introduced by augmentation did not provide additional generalization benefit for this dataset. This may reflect the fact that class-weighted learning alone was sufficient to address class imbalance, while the introduced spectral variability may have partially weakened subtle class-discriminative Raman patterns. This is consistent with previous spectroscopy studies showing that augmentation-based or imbalance-handling strategies do not always yield statistically significant improvement, particularly when class similarity is high or the generated variability is not sufficiently representative of the underlying spectral distribution [34]. Although conventional chemometric methods such as PLS-DA and PCA-DA remain competitive [35, 36] and interpretable for Raman classification, particularly on relatively small datasets, the proposed 1D-CNN was explored here to capture broader multivariate and potentially nonlinear spectral relationships in an end-to-end manner [37, 38]. In this context, the CNN was employed to model full-spectrum biochemical variation that may not be fully represented by isolated peak-wise analysis alone.    

The comparative stability and reproducibility of serum as a diagnostic biofluid further strengthened the reliability of Raman-based classification. Serum was selected as the analytical biofluid because it provides a clinically relevant matrix for detecting systemic biochemical changes, while being less influenced by local oral-environment factors than saliva [39]. The observed spectral distinctions, particularly in the amide, CH-bending, and (1300–1350 cm ), egions, are consistent with reported Raman changes associated with alterations in proteins, lipids, and nucleic-acid-related components during malignant transformation [40, 41]. Additionally, validation using a temporally separated dataset—collected from new patients under identical acquisition protocols—confirmed the model’s applicability to real-world clinical scenarios. The consistent accuracy obtained from this independent validation supports the generalizability and reliability of the proposed approach. Overall, the study demonstrates that coupling Raman spectroscopy with a tailored 1D-CNN framework can provide a non-invasive, rapid, and reproducible diagnostic tool for early detection and classification of oral cancer. The model’s interpretability, balanced class performance, and robustness to data variability establish a strong foundation for future large-scale clinical translation and cross-site validation.

A limitation of the present study is the relatively small number of analyzed patients, which may restrict the statistical power and generalizability of the findings. Although this work represents a proof-of-concept study demonstrating the feasibility of serum Raman spectroscopy combined with 1D-CNN analysis for three-class oral lesion classification. The results are encouraging, the limited sample size and single-center design mean that broader clinical applicability should be interpreted cautiously and requires validation in larger multi-center cohorts.

Conclusion

The proposed 1D-CNN model demonstrated the feasibility of classifying serum Raman spectra into normal, precancer, and cancer groups. The final selected non-augmented model achieved an overall accuracy of 83%, with macro-precision, macro-recall, and macro-F1 values of 85%, 84%, and 83%, respectively, indicating balanced three-class performance. Consistent results on the independent test cohort further supported the generalizability of the framework within the same clinical environment. These findings highlight the promise of combining serum Raman spectroscopy with deep learning as a non-invasive strategy for oral cancer-related classification, although larger multi-center studies are still needed for broader clinical validation.

Author contributions

Mukta Sharma: Writing – original draft, Validation, Methodology, Data curation, Conceptualization, Software and Formal analysis and Shiang-Fu Huang: Writing – review & editing, Supervision, Resources.

Funding

This work was partially supported by Grant NSTC 113–2222-E-167–003 -MY2 from the National Science and Technology Council (Taiwan).

Data availability

Data will be made available from the corresponding author upon request.

Declarations

Ethical approval and consent to participate

This study was performed in line with the principles of the Declaration of Helsinki. Ethical approval was granted by the Institutional Review Board of the Chang Gung Medical Foundation, Taiwan (IRB Nos. 202300682B0 and 201800439B0). Written informed consent was obtained from all participants included in the study.

Consent for publication

Written informed consent was obtained from all participants for the publication of their clinical data. All patient information was systematically deidentified prior to analysis to ensure confidentiality.

Competing interests

The authors declare no competing interests.