An infrared spectroscopic blood test for non-small cell lung carcinoma and subtyping into pulmonary squamous cell carcinoma or adenocarcinoma

2.1. Patient collective

In close collaboration between the Ruhrlandklinik, the Institute of Pathology, and the PURE institutes, strictly defined standard operating procedures (SOP) for tissue and blood sampling according to the rules of Good Epidemiological Practice were developed. Following these protocols, patients were fully informed about the study and gave their written consent. Epidemiologic data were collected along with blood samples among patients of the Ruhrlandklinik Essen from August 2013 to February 2015. Blood was processed to serum, EDTA-, and tri-sodium citrate stabilized plasma with clinical routine equipment (Sarstedt AG&Co., Nümbrecht, Germany) obeying strict SOPs. The samples were frozen within less than 60 min after sampling. All samples were stored at 80°C until experimental use. The established diagnosis of bronchoscopy and histopathology served as gold standard. This study complies with the applicable ethical guidelines and was approved by the Ethics Committee of the Ruhr-Universität Bochum (Ethics vote 13-5420-BO) and the Ethics Committee of the University Duisburg-Essen, Essen, Germany.

NSCLC was the most abundant class of malignant diseases of the entire study (81%). Squamous-cell carcinoma (SCC) and adenocarcinoma (AC) patients represented the most abundant subtypes of NSCLC (together 79%). Consequently, three most populated, most homogeneous patient classes were assembled: (i) a cancer-free disease control group without confirmed initial cancer suspicion (53 patients), (ii) 54 primary SCC patients and (iii) 54 patients suffering from primary AC (Table 1).

The first recruited 19 SCC patients and the first 34 AC patients were assembled to a general ‘lung carcinoma’ class (LC). The ratio equals the abundance of these NSCLC tumor subtypes in the performed study.

Patients were recruited successively upon admission into the hospital. Thus, the selection of the first patients of a group equals a random selection among the participants. A selection of equally powered groups was necessary to prevent classifier distortion [21,22].

The group of control patients reflects the collectives of patients with non malignant diseases of the thoracic cavity presenting to the Ruhrlandklinik Essen during recruitment. The group consisted of patients with pneumonia, pleuritis and other inflammatory changes ( n = 34 ), interstitial lung diseases ( n = 7 ) and benign tumors or lesions ( n = 6 ). Multiple pathologies were observed in several patients. In six patients, malignancies were excluded, but a definitive diagnosis could not yet be obtained.

Generally, patients with lipemic or hemolytic blood samples were excluded from the current analysis, as well as those with lung metastases of other tumors. Gender and age distributions were similar among the patient classes (Table 1). Tumor grading was matching well among the groups. Spectral differences of tobacco consumption within the patient groups (smoker/non-smoker/never smoker) were researched with the same pattern detection algorithms as described in 2.3–2.5, but not detected in the infrared spectra of blood samples. Therefore, confounding contributions of gender, age, tumor grading and tobacco consumption to the data are not expected. All patients were recruited before obtaining the diagnosis. Therefore, it is unlikely that confounding dietary habits were unique to one patient class.

For quantitative normalization, 3 µl of 500 mM aqueous KSCN solution were added to each 47 µl serum, EDTA and citrate plasma before analysis in technical quadruplicates. The automated HT-FTIR-measurements (Vertex 70v FTIR spectrometer, HTS-XT extension, Twister robotic plate feeder, Bruker Optics GmbH, Ettlingen, Germany) of robotically spotted samples (50 nl per well) (instrumentTwo, M2 Automation GmbH, Berlin, Germany) on 384 well silicon MTPs (Bruker) were performed as described with 4 cm⁻¹ resolution [21,22].

2.3. Data processing

The automated, self-parameterizing or pre-parameterized procedures for scaled subtraction of atmospheric water vapour, smoothing, adaptive iteratively penalized least squares (airPLS) baseline correction, outlier removal, averaging normalization, differentiation and combination of spectra were described in detail [21].

To obtain a representative spectral vector for each patient, the absorbance, 1st and 2nd derivative absorbance spectra of all three blood preparations were concatenated. Hence, we obtained vectors of 5751 features with 2 cm⁻¹ spacing for each patient.

2.4. Feature selection

Crucial to such an analysis is the reduction of classification problem dimensionality by removal of redundant and uncorrelated information. With the most discriminative data, a best possible class separation should be achievable, with a reduced chance of overfitting the data. Two entirely different algorithms were applied. One computationally efficient algorithm with low requirements for processing power is termed maximum relevance, minimum redundancy (MRMR). It is based on data discretization and z-testing [23] and was successfully applied to nonlinear multivariate classification problems [6,21–23]. The basic algorithm can be downloaded from http://www.mathworks.com/matlabcentral/fileexchange/ 14,916 (October 16, 2015). The 100 most discriminative features were determined for each dataset, and ranked by the algorithm. A consecutive step-down selection for highest accuracy in 1000 independent leave-one-third-out MCCV with linear discriminant analysis (LDA) classifiers was performed. Supplementing the single top ranked feature, the lesser ranked up to 100 features were included into the calculation. The highest-ranked feature set performing with highest average accuracy in 1000 fold leave-1/3rd-out MCCV was identified as the MRMR selection result.

An alternative wrapping approach with iterative random forest classification was performed, demanding high processing power. It was successfully applied before and described in detail [21,22]. Briefly, random forests were used to determine the Gini-importance of single spectral features for correct classification [8,10,17,18]. After repeat selection from MC derived data subsets comprising 90% of the total dataset, the selection frequency of each identified feature was mapped. Based on the average accuracy, the best predicting feature sets were registered into a pool of selected features. This pool was analyzed by stepping down in search of a minimum selection frequency threshold. For each threshold, the identified feature sets were individually validated for optimum average accuracy in a 1000 fold LDA leave-one-third-out MCCV. The best performing feature set determined the minimum selection frequency threshold, which is given in Table 2.

The identified classification-characteristic features were checked for agreement with spectral contributions of the silicon substrate in absorbance, 1st and 2nd derivative spectra. Only one identified feature (1293 cm⁻¹ in the 2nd derivative serum component) coincided with a broader substrate absorbance band at 1296 cm⁻¹. MCCV indicates this feature as redundant. This is discussed in the result section.

A dual cross validation concept with two routines of proven performance [22] was followed. As a compromise to the limited set of patients in the study, a strict leave-one-third-out MCCV scheme was applied to all validation procedures. Classifiers were trained on a randomly selected 2/3 patients of the dataset to predict the left-out 1/3 subjects.

An algorithm with low time complexity is the classifying linear discriminant analysis (LDA). The Matlab provided routine ‘classify’, with à priori class membership estimation and a linear discriminant function was used. For robust results, a 1000 fold MCCV scheme was maintained with LDA classifiers.

A complex ensemble random forest classifier was applied as the demanding alternative in regard to processing power and classifier complexity [21]. In a random forest, prediction is achieved by majority vote of the included decision trees. Here, the prediction was deduced from the majority vote of an ensemble of 1001 random forests.

Following standard nomenclature, the percentage of correct classifications was defined as accuracy, whereas the percentage of true positive predictions among all positives reflected the sensitivity. Specificity is given as the percentage of true negative among all negative predictions.

2.6. Bioinformatics environment

Random forest routine calculations were performed within the Matlab environment, version 2012a with the R-project based [15] Matlab port (downloadable from http://code.google.com/p/randomforest-matlab/, October 21, 2015) on a High-Performance Computing Server Supermicro SYS-5086B with 8x Intel^® Xeon^® Westmere EX (E7-8837, 2.66 GHz, 8-Core), 512 GB RAM. MCCV and MRMR feature selection were performed on a standard office PC equipped with Intel Core2Quad CPU Q9650@3.0 GHz, 8 GB RAM running Matlab 2012a.

3.1. The spectral dataset

Infrared absorbance spectra were obtained from 483 blood samples. 1932 spectra of the quadruplicate analysis were condensed and concatenated to representative spectral vectors of 161 patients, of citrate plasma, EDTA plasma, and serum (Fig. 2(A), (B), (C)). This resulted in 5751 features per patient. Spectra were normalized with the C–N triple bond absorbance band between 2150–2000 cm⁻¹ of KSCN spiked to the sample. This maintained the absorbance band intensities as qualitative markers for the concentration of included substances, as e.g. total protein, total lipids, phosphates and glucose (Fig. 2, regions (i), (ii)). Due to the natural variation of concentrations, a clear assignment by visual inspection of marker bands is futile in the absorbance, 1st (regions (iii), (iv)) and 2nd derivative (regions (iii), (iv)). Sharp difference bands at 1245 cm⁻¹ and 1182 cm⁻¹ indicate spectral contributions of the silicon substrate to some spectra. This was likely caused by alteration of the substrate surface between sample-free reference measurement and the analysis of the deposited sample spot. Although not yet clearly evaluated, a repeated plasma treatment may have sensibilized the silicon surface for reactivity with atmospheric water. These bands are not expected to provide classification relevant information. However, the selected features were thoroughly checked to not coincide with these bands and the respective side lobes after derivatization. Some presented features appear close (Table 3), but are positioned next to bands or close to neutral turning points between sidelobes. One feature (serum 2nd derivative, Fig. 2(C) (vi), 1293 cm⁻¹) close to one such band at 1296 cm⁻¹ was selected by the MRMR algorithm. This feature is nevertheless redundant for classification, as will be shown later.

OPEN IN VIEWER

Fig. 2.

Absorbance spectra, (i, ii), 1st derivative (iii, iv), and 2nd derivative spectra (v, vi) of citrate plasma (A), EDTA plasma (B), and serum samples (C) as recorded and calculated from disease control (black), SCC (blue) and adenocarcinoma patients (red). The analysis was constrained to the respective C–H-stretching ((i), (iii), (v): 2800–3200 cm⁻¹) and the fingerprint regions ((ii), (iv), (vi): 1750–875 cm⁻¹).

An approximate assignment of vibrational bands to functional groups, and to abundant blood substances featuring those groups is given in Table 4. These assignments are not unique, because FTIR absorbance bands are not unique to single substances contained in blood. For example, along with proteins, urea as well features strong absorbance bands in the 1700–1500 cm⁻¹ region, due to the contained C=O and NH₂ groups. Therefore, a unique scheme of increased and diminished substances in cancer-patient blood cannot be developed based on the identified infrared spectral feature patterns.

First, to demonstrate the basic separability of NSCLC patients from the disease control group based on blood samples. Common spectral features of squamous cell carcinoma and adenocarcinoma were identified and verified in MCCV.

The lung cancer class, consisting of 19 SCC and 34 AC randomly selected patient samples, was separable from the disease controls with both feature sets. The projection of an LDA classifier into the score space of the first components of a principal component analysis (PCA) indicates a correct separation of more than 50% of patients (Fig. 3(A)). This is confirmed with MCCV with both LDA and RF classifiers on the dataset, indicating 71% and 72% accuracy, respectively (Fig. 3(B), Table 2). Respective sensitivities of 67% and 71%, and specificities of 75% and 73% were achieved. The RF based selection offers a visibly better class separation in the PCA projection (Fig. 3(C)). Although both MCCVs indicate similar accuracies of 71% and 72%, they also indicate improved sensitivities of 70% and 75% with specificities of 72 and 69%, respectively (Fig. 3(D), Table 2). The MRMR algorithm identified 12 relevant spectral features, whereas only two passed the rigorous RF based algorithm (Table 3). Although this appears as a rigorous feature reduction, both the PCA analysis (Fig. 3(C)) and the MCCV results (Fig. 3(D)) justify the feature selection. In this special case, the PCA-score plot illustrates the LDA separation of both classes by the intensities of the EDTA plasma absorbance feature at 1261 cm⁻¹ (abscissa) and at 941 cm⁻¹ of the 2nd derivative of the serum absorbance spectra (ordinate).

OPEN IN VIEWER

Fig. 3.

Performance of lung carcinoma spectral features extracted with the MRMR (A, B) and the RF method (C, D) versus control patients, determined with LDA and RF classifiers. The PCA score plots (A, C) visualize a qualitative separability, whereas the quotient ratios (B, D) quantify average MCCV results with accuracies above 70%.

The disease control group represents a typical variety of cases as expectable in a clinical setting. Therefore, it is an inhomogeneous group dominated by inflammatory lung diseases, and some less abundant non-inflammatory cases. Thereby, our classification system was tested with realistic control patients. It is likely attributable to the inhomogeneity of the current patient groups, that the accuracy of cancer detection is decreased in comparison with our earlier study on bladder cancer [22]. That study achieved an average accuracy of up to 92%, but was performed with the less inhomogeneous patient groups of urothelial carcinoma versus (severe) urinary tract inflammation.

To pursue subtyping of NSCLC, spectral indicators of 54 primary SCC patient blood samples were identified and verified against 53 cancer-free control patients.

Both feature selection algorithms identified indicative spectral features, which all cross-validated with a significantly higher average accuracy than 50%. Four MRMR selected features indicated a class separation in the score plot of the first two PCAs (Fig. 4(A)). MCCV resulted in respective accuracies of 68 and 66%, sensitivities of 66 and 68%, and specificities of 69 and 64%, with LDA and RF classifier MCCV (Fig. 4(B), Table 2). Although spectral contributions of the silicon substrate (Si–H, Si–O, Si–O–H, C–Si–O and local surface defects) are usually avoided by referencing a sample position with the identical substrate before sample application, one MRMR selected feature at 1293 cm⁻¹ coincided with an identified artefact at 1296 cm⁻¹ on background noise level intensity. MCCV were repeated without this feature. LDA MCCV resulted in 66 ± 7 % accuracy, with a sensitivity of 64 ± 11 % and a specificity of 69 ± 11 % . RF MCCV resulted in 68 ± 7 % accuracy, with a sensitivity of 69 ± 10 % and a specificity of 67 ± 12 % . That means, the LDA MCCV accuracy was reduced by 2% units, whereas the RF MCCV determined accuracy gained 2% units by not considering the 1293 cm⁻¹ feature (Tables 2 and 3). It appears not essential to the classification and could therefore represent a borderline redundant feature.

OPEN IN VIEWER

Fig. 4.

Performance of squamous cell carcinoma spectral features extracted with the MRMR (A, B) and the RF method (C, D) versus control patients, determined with LDA and RF classifiers. The PCA score plots (A, C) visualize a qualitative separability, whereas the quotient ratios (B, D) quantify average MCCV results.

A significantly better classifier performance was obtained with 11 RF selected features. The two dimensional projection of the PCA score plot indicates a specific, but not necessarily sensitive class separation (Fig. 4(C)). Quantitative numbers of the MCCV analysis do not confirm this impression: LDA and RF indicate an average accuracy of 73%, sensitivities of 71% and 74%, and specificities of 75% and 71%. Thus, very similar results were obtained using these entirely different classifiers (Fig. 4(D), Table 2).

Consequently, the spectral discrimination of disease control patients from those suffering from primary adenocarcinoma was researched. To maintain balanced class sizes, 53 patients of each DC and AC were compared.

OPEN IN VIEWER

Fig. 5.

Performance of adenocarcinoma spectral features extracted with the MRMR (A, B) and the RF method (C, D) versus control patients, determined with LDA and RF classifiers. The PCA score plots (A, C) visualize a qualitative separability, whereas the quotient ratios (B, D) quantify average MCCV results.

The PCA score plot of 34 MRMR selected features indicates an essential class separability with approximately equal numbers of false positive and false negative assignments (Fig. 5(A)). MCCV confirms the qualitative impression: repeated LDA MCCV resulted in an average accuracy of 69%, a sensitivity of 68% and a specificity of 70%. With RF MCCV, significantly higher values were obtained: 77%, 77% and 78% (Fig. 5(B), Table 2). The seven RF selected features, on the other hand, indicated a good separability in the PCA analysis (Fig. 5(C)). Both LDA and RF MCCV resulted in higher average numbers. With LDA, an average accuracy of 75%, sensitivity of 72% and specificity of 77% was achieved. The RF MCCV resulted in an accuracy of 79%, with a sensitivity of 77% and a specificity of 80%. Thus, AC patients are clearly separable from DC patients based on the infrared spectral blood analysis.

Finally, the separability of SCC from AC patients in regard to NSCLC subtyping was analyzed. Thus, 54 patients of each class were analyzed.

Again, both MRMR and RF algorithms identified differentiating feature sets. Ten MRMR selected features roughly discriminated the two disease classes in the PCA score projection (Fig. 6(A)). Repeated MCCV indicates a relatively high separation quality. Average accuracies of 73% and 75%, sensitivities of 75% and 74%, and specificities of 71% and 76% were achieved, with respective LDA and RF MCCV (Fig. 6(B), Table 2). Contrastingly, the RF algorithm selected 29 features (Table 3). Using those, the PCA analysis indicated an improved class separability (Fig. 6(C)). Further, the highest cross validation results were achieved. The LDA and RF MCCVs resulted in respective average accuracies of 80% and 79%, sensitivities of 82% and 80%, and specificities of 77% and 79% (Fig. 6(D), Table 2).

OPEN IN VIEWER

Fig. 6.

Performance of spectral features extracted with the MRMR (A, B) and the RF method (C, D) to differentiate SCC from adenocarcinoma patients, determined with LDA and RF classifiers. The PCA score plots (A, C) visualize a qualitative separability, whereas the quotient ratios (B, D) quantify average MCCV results.

Summing up, the presented FTIR spectroscopic DPR approach performed similarly well as a recently reported multi-marker panel approach for screening of oncologic diseases [29]. In a 12 year study on 41,516 patients, Wen et al. achieved a sensitivity of 75% for lung cancer: 27 of 36 lung cancer patients were correctly identified with an eight molecule marker-panel. Another very applicable strategy for cancer detection by blood analysis was termed liquid biopsy, meaning the identification and characterization of cell-free DNA (cfDNA) [25] or circulating tumour cells (CTC) [30]. Sensitivities ranging from 24–100% were reported [30]. As a 73–79% accuracy is clearly insufficient for screening, we propose to evaluate the combination of multi marker panels and a liquid biopsy assay with the DPR approach. Such a testing scheme is likely to yield improved results. Further, the scheme could be shaped for either optimum sensitivity or optimum specificity, exploiting the strength of each single technique.