Urine based near-infrared spectroscopy analysis reveals a noninvasive and convenient diagnosis method for cancers: a pilot study

Participants and urine sample collection

A total of 181 cancer patients as well as 146 healthy controls were recruited from Zhejiang cancer hospital, with the approval of the Ethics Committee of Zhejiang Cancer Hospital (IRB-2023-375). We received written informed consent from all the participants. Of 181 cancers, there were 62 lung cancer cases, 32 gastric cancer cases, 19 cervical cancer cases, 14 colon cancer cases, 12 thyroid cancer cases, 11 ovarian cancer cases, 10 breast cancer cases, nine liver cancer cases, nine nasopharyngeal carcinoma cases, two bladder cancer cases, and 1 kidney cancer case. The age and sex between cancer group and healthy control had no significant difference (Table 1). Before collecting a urine specimen, one should avoid vigorous exercise. In this study, morning urine was collected as the specimen, which refers to the first midstream urine sample after waking up in the morning. During collection, be sure to avoid contamination from urethral secretions, menstrual blood or vaginal secretions, semen or prostatic fluid, stool, and other substances. Use a disposable urine cup for the urine specimen, and label the container with the patient’s name and unique identifier. The urine specimen should be sent for testing within 2 h after collection. Upon receiving the specimen, mix it well, draw one mL of urine, and store it at −80 °C until analysis. The study was approved by Zhejiang Cancer Hospital Committee and informed consent was received from all patients.

NIRS data collection

Frozen urine samples were thawed at ice before NIRS analysis. The procedure for NIRS analysis was according to our recent plasma based NIRS study (Zhu et al., 2023). Briefly, Antaris™ II FT-NIR analyzer (Thermo Fisher Scientific, Waltham, MA, USA) was used for NIRS collection with air as the reference. A total of 200 µL urine sample was loaded in a quartz colorimetric tube with an optical path of six mm, and NIRS was generated by averaging 32 successive scans, ranging from 4,000 to 10,000 cm⁻¹, with a resolution of four cm⁻¹. The spectra were measured based on molar absorptivity, and the average spectrum was obtained for each sample by taking three measurements and processing the data using TQ Analyst 8.0 software.

Data analysis

Raw NIRS data and pretreated NIRS data

Raw NIRS data is listed in File S2. Each NIRS spectrum consisted of 1,557 points ranging from 3,999.64 to 10,001.03 cm⁻¹. The average NIRS of cancers almost overlapped with that of healthy controls (Figs. 1A–1B).

PCA analysis showed no separation trend between cancers and controls

When PCA analysis of raw NIRS data from train dataset without scaling, the first component already had 100% of the data variance, but there was no significant separation between cancers and controls (Fig. 1C). The PCA analysis of NIRS data scaling with “center” and “scale” also showed that there were greatly overlapping between cancers and control (Fig. 1D). The result implied that urine NIRS data had no direct linear relationship with their classes.

Influence of pretreatment methods on the prediction performance of PLS and SVM

For PLS modeling with different pretreated NIRS data, the second derivative pretreatment improved the prediction accuracy for test samples, while the other methods did not influence the prediction performance or even slightly decreased the prediction accuracy, such as MSC and the second derivative, compared to modelling with raw NIRS data (Table 2 and Figs. 2A–2F). When further combined with scaling treatment, the result showed that the combinations of scaling and MSC, the first derivative, and the second derivative, evidently increased the prediction performance (Table 2 and Figs. 3A–3F). Thus, PLS modeling with NIRS after both the second derivative and scaling achieved the best prediction performance, in which the prediction accuracy in test set was 0.78, and the AUC _ROC was 0.853). The detailed prediction performance for modeling with other treatments were demonstrated in Table 2, and Figs. 2 and 3.

For SVM modeling with single pretreatment, the result showed that SVM with the second derivative pretreatment achieved the best prediction performance, and SG smoothing had no influence on the prediction accuracy, while other methods decreased the prediction performance (Table 2, Fig. 4). In modeling with combination of scaling and other pretreatments, the result showed scaling dramatically improved the prediction performance for models. The SVM models with combinations of the first derivative pretreatment and scaling had the highest accuracy of 0.853. In terms of ROC analysis result, SVM models with combinations of raw and scaling, SG and scaling reached the highest AUC _ROC of 0.927. Therefore, SVM with scaling treatment was the best models in this study. The detailed prediction performance for SVM modeling were illustrated in Figs. 4 and 5, and Table 2. Additionally, the cost value in SVM was optimized at 8, with which the model was able to reach its max prediction accuracy of 0.862.

In our study, SG pretreatment had almost none influence on the prediction performance for both PLS and SVM models, leading to the same prediction results and ROC curves from modeling with raw data and SG pretreated data (Table 2, Figs. 2A, 2B, 3A, 3B, 4A, 4B, and 5A, 5B). During modeling, it was observed that low specificity of PLS and SVM was the main factor resulting in a low prediction outcome. The different pretreatment methods influenced the specificities of models. For example, the first and second derivative decreased the specificities in SVM modeling, while scaling NIRS significantly increased the specificities in SVM models (Table 2). The detailed result of prediction performance in training steps were listed in File S3.

Modeling importance of NIRS feature in SVM model

SVM-RFE algorithm ranked 1,557 NIRS features, which was shown in File S4. By progressively modeling with a combination of top 262 features in SVM, the overall prediction accuracy reached its max value of 0.862 (Fig. 6A). Besides, SVM model had its max sensitivity of 0.908 with only top 1 feature, and had its max specificity of 0.909 with top 220 features (Figs. 6B, 6C). Among the top 262 NIRS feature, 201 were in Band II (8,500 ∼5,500 cm⁻¹), which referred to 1st overtone and their combinations of CH2/CH3/OH/NH stretching), eight were in Band I (12,500 ∼8,500 cm⁻¹), which referred to 2nd overtone and their combinations of CH2/CH3/OH/NH stretching, and 53 were in Band III (5,500 ∼4,000 cm⁻¹), which referred to combinations of CH2/CH3/OH/NH stretching and 2nd overtone of C =O stretching (File S5).

Advantages of urine based NIRS analysis for cancer diagnosis

Compared to the traditional cancer diagnosis methods, such as carcinoembryonic antigen test, CT/MRI scans, and pathology, which would cause pain or harm to some extent, urine based NIRS analysis has obvious advantages of noninvasiveness, convenience, and low cost. Therefore, urine based NIRS has a great potential as a cancer diagnosis application, especially in a daily monitoring way. Though LC-MS based omics are widely used in cancer biomarker discover studies, the equipment and analysis cost are expensive, and professional and technical personnel are required to perform the analysis. In contrast, NIRS equipment and its analysis cost is much lower than LC-MS, and it needs only a very simple sample preparation or even without preparation. Because NIRS equipment is relatively small, cheap, and portable, a further picture that this pilot study hints at are that there will be NIRS detectors in the home for daily monitoring for healthy status, for example, a “smart toilet” with a NIRS detector. Given these advantages of urine based NIRS diagnostics, further large-scale multicenter studies are encouraged.

Machine learning algorithms

Unlike genomics, proteomics, and metabolomics, which can identify molecules in biological samples, NIRS provides comprehensive indicators of the functional groups of molecules. Machine learning algorithms are necessary to be combined in NIRS data analysis for diagnosis. There are a variety of machine learning algorithms available for modeling with NIRS data. PLS is one of the most commonly used algorithms in NIRS field due to its advantages as follows: (1) PLS is suitable for situations where the number of samples is less than the number of variables; (2) PLS provides interpretable results; (3) PLS can handle multicollinearity. While PLS is a linear model, and it was not always the best algorithm for complicated relationship between variable and interests. SVM algorithm (Normanno et al., 2018) could be used where the relationship between spectra and interest is presented as a non-linear. In this project, though PLS took only about half of the computational time of SVM, the prediction accuracy was much lower than that of SVM. In consistent with this, our previous NIRS based cancer diagnosis projects demonstrated SVM was more suitable than PLS in modeling with NIRS data from liquid biopsy samples.

Pretreatment is very important in NIRS data modeling, and this study illustrated that different models need their specific pretreatment methods, it is crucial to do a proper investigation in pretreatments. This project showed that only the second derivative improved the PLS, while the “scaling” and “center” pretreatment significantly improved the SVM model. Other pretreatments, including SG smoothing, multiplicative scatter correction, baseline removal, the first derivative, played a limited role in model performance, and even decreased the prediction accuracy in specific model, such as SVM modelling with the first derivative pretreated NIRS data. Therefore, it is highly recommended to carefully select pretreatment for the specific algorithm.

Assignment of the NIRS features

Though it is difficult to assign these NIRS features directly to specific molecules, due to weak absorbance, wide peak width, and overlapping of the NIRS peaks, NIRS can still provide a lot of useful information about the functional groups of the samples. There are three main NIRS bands, including Band I, Band II, and Band III, which cover a variety of functional groups of biological molecules, including metabolites, DNA, and proteins. In this study, among the top 262 NIRS features of SVM model, 201 were in Band II, which contained water spectral peaks. In the past, the main disadvantage of NIRS in liquid biopsy analysis was from influence from water spectral peaks in this range. Since Dr. Roumiana Tsenkova established a theory- “Aquaphotomics”, in which water band was the main study object and the compounds in water interact with the OH of water, and then change the water spectral pattern (Tsenkova, Kovacs & Kubota, 2015; Tsenkova et al., 2018). With the inspiration of “Aquaphotomics”, water status might be changed due to changes in concentrations of compounds in urine, including metabolites, DNA and proteins, which are widely reported to associated with diseases by metabolomics, proteomics, and genomics. Indeed, we tried modeling with a specific water spectral area, but we failed to obtain a better prediction outcome than that with the whole spectra (prediction accuracy in testing: 0.761). However, it is worthy to investigate the potential diagnostic role of “Aquaphotomics” in urine-based NIRS diagnosis. However, combinations with other analytical tools, such as LC-MS and NMR, are still needed to identify the specific biomarkers for diseases.

Limitation is sample size

This is only a pilot study, the limitations in the following aspects must be listed. First, the sample size is relatively small, and urine samples were from only one center. Thus, our established urine based NIRS diagnosis method should undergo further validation with a large sample size study, especially containing validation samples from other centers. Second, this pilot study had too many types in a small size cancer group, and it might lead to a bigger intragroup variance and then a lower prediction accuracy. Thus, further study with a proper number of cancer types from some specific system, and each cancer type has adequate cases for modeling. Third, the algorithms applied in this project were only simple ones, while more complicated model, such as deep learning algorithm, has not been used due to the limited sample size.