In Figure 1, the entire pipeline implemented for this study is presented. In the first step, labeled data were generated by acquiring drainage and urine samples and performing spectral measurements as well as laboratory analyses on them. For each collected sample, 3 spectra and a series of laboratory markers were obtained. This labeled data was used to train AI models for each of the laboratory markers (urine and drainage). In the second step of the pipeline, the spectra were preprocessed in preparation for the AI models. This step involved normalizing the spectra. For the CNN model, an additional step was performed by converting the spectra to 3-channel images. The labeled data was eventually fed to 3 separate AI models: a CNN, a PLS-DA, and an RF classification model.

This study was approved by the Ethics Committee of the University Hospital Essen (21-10402-BO). Informed consent was obtained from all patients. The data included in this study were fully anonymized. Participants received no compensation.

The dataset used in the present work consists of 454 (181 patients) drainage and 401 (168 patients) urine samples. The samples were acquired from patients aged between 0 and 85 years at the University Hospital Essen. The age distribution of the drainage dataset has a median and IQR of 57 and 22, respectively. For the urine data, the median amounts to 56 and the IQR to 20.25.

Each sample was divided into 2 batches. The samples were then frozen and stored at −80°C until further processing. All samples from the first batch were analyzed at the central laboratory of the University Hospital Essen. A total of 14 drainage markers (Table 1) and a total of 11 markers for the urine samples (Table 2) were examined. For each sample, the measured markers were binarized using predefined cutoff values. If a specific marker was higher than the corresponding cutoff value, the marker was labeled as pathologic. Otherwise, it was labeled as healthy. Those cutoff values were defined by the guidelines provided by the central laboratory (version 1.4 dated September 21, 2021). As there are no predefined cutoff values for drainage fluids, standard values for serum were used as a reference for those samples. However, for the markers hemoglobin and erythrocytes, the cutoff value for pathology was set to 0 because their presence in surgical drain fluids is universally considered pathological and may even indicate relevant postoperative bleeding [25].

Two exceptions to the binarization procedure were made for the majority of urine markers. Most of these were measured using urine dipstick tests, which estimate the concentration of the marker using a categorical scale. A measurement was marked with the minus symbol “−” if the concentration of the marker is not high enough to be detectable. Pathologic concentrations of the marker were marked with a series of plus symbols “+.” Therefore, a marker was classified as pathologic if the laboratory stick measurement showed at least one “+” symbol. Otherwise, the marker was classified as healthy. The second exception to the binarization procedure was made for pH. This marker does not present a single cutoff value but rather a range of values for which the marker was considered normal and was, therefore, considered healthy. Values measured outside the defined ranges were defined as pathological.

Additionally, in Tables 1 and 2, the ratio between the minority and majority classes of each binarized drainage and urine marker is listed. This value provides important information on the balance between pathological (red bars) and healthy (green bars) samples within a specific fluid marker. Therefore, a perfectly balanced dataset where both classes include the same number of samples would produce a ratio of 1. On the contrary, a dataset where 1 of the 2 classes does not contain any sample would produce a ratio of 0. As shown in the results, this value greatly affects the quality and performance of the trained models.

For each sample in the second batch, spectral measurements were performed using a compact mini-spectrometer. The mini-spectrometer was integrated with a self-developed lens array and an electrical current- and temperature-controlled hyperspectral illumination source. This setup ensures broad-spectrum illumination of the samples. The evaluation platform, although relatively large in size, was designed to accommodate the assessment of multiple illumination angles simultaneously and did not yet focus on size reduction during its development. A schematic representation of the spectrometer is presented in Figure 2.

The measured spectra consist of 288 data points captured between 313.08 and 874.27 nm. The data collection was improved by illuminating each sample from 3 different angles: direct transmission (DT), angular transmission (AT), and angular reflection (AR). The exposure time for each angle was fine-tuned to achieve an optimal signal-to-noise ratio, amounting to 20, 200, and 320 µs for the settings DT, AT, and AR, respectively. Therefore, the input data for the AI models can be thought of as a feature matrix with the shape of (N, 3, 288). Where N is the number of samples of a specific fluid marker, 3 is the number of measured spectra per sample (DT, AT, and AR), and 288 is the number of datapoints measured for each spectrum. An example measurement of a drainage sample is shown in Figure 3.

Following each measurement, the sample was discarded, and the evaluation platform’s tubing was thoroughly flushed with sterile water to prevent cross-contamination between samples. After completing the batch measurement, the tubing was flushed with isopropyl alcohol followed by water to prevent sample cross-contamination and bacterial growth.

For each fluid marker dataset, the samples were grouped at the patient level to ensure that all measurements from a given patient were kept together. Thereafter, a patient-wise train-test split with an 80/20 ratio was performed. This approach prevents data leakage, ensuring that all samples measured on a specific patient are not spread across multiple datasets [27,28]. Consequently, the training set was further subdivided into 3 nonoverlapping subsets, grouped at the patient level and stratified with respect to the target label (eg, ratio between pathological and healthy subjects). Those datasets were used for cross-validation training. Each fold was then tested on the test set, and the classification probabilities were averaged across all 3 folds to obtain the final classification probabilities on the test data.

Due to the highly imbalanced datasets of many fluid markers, the Matthews correlation coefficient (MCC) was chosen as the primary metric to track and maximize. Alongside MCC, the F₁-score and the true positive rate (TPR) were measured. Additionally, the receiving operating characteristic (ROC) curve was computed together with the corresponding area under the curve (ROCAUC) of each model. Additionally, the predictability of the trained models was assessed by comparing them to a dummy classifier.

In Table 3, an overview of the performance of the AI models trained on all drainage markers is presented. For each of the markers, hemoglobin, bilirubin, albumin, lactate dehydrogenase, total protein, and mononuclear cells, at least 1 model was trained for which the MCC score measured on the test dataset was higher than 0.5. For those markers, F₁-scores of 0.71 or higher were measured for at least 1 of the AI models. Except for mononuclear cell models, the ROCAUCs of the mentioned markers were consistently higher than 0.8 on all 3 AI approaches. To assess the overall performance of the 3 AI approaches, the average value of the MCC scores across all fluid markers was computed. The PLS-DA and RF methods yield comparable mean MCC scores of 0.37 (SD 0.28) and 0.37 (SD 0.26), respectively, whereas the CNN approach scores slightly higher at 0.4 (SD 0.3). A Friedman test showed no statistically significant difference in performance between the models (χ²₂=2.31, P=.315, Kendall W=0.08).

The models trained on hemoglobin and bilirubin data stand out from the remaining markers, showing the highest performance measured in this study. The CNN approach performed best for hemoglobin data with an MCC score of 0.83, F₁-score of 0.96, ROCAUC of 0.97, and a TPR of 0.99. The CNN approach delivered promising results also on bilirubin data: MCC score of 0.81, F₁-score of 0.85, ROCAUC of 0.95, and a TPR of 0.74.

In Figure 6, the confusion matrices for drainage bilirubin and hemoglobin classification on the test dataset are shown, along with the ROC curves calculated from the predicted classification probabilities. The confusion matrices and ROC curves of the other drainage and urine marker models can be inspected in Figures S2 and S3 in Multimedia Appendix 1.

The models trained on urine data exhibit overall lower performance when evaluated on the test datasets. The mean MCC scores measured across all markers amount to 0.29 (SD 0.26; PLS-DA), 0.28 (SD 0.29; RF), and 0.31 (SD 0.27; CNN). A Friedman test showed no statistically significant difference in performance between the models, with a small to moderate effect size (χ²₂=5.15, P=.08, Kendall W=0.23). Nevertheless, the markers bilirubin, erythrocytes (hemoglobin), protein, urobilinogen, and albumin were able to achieve MCC scores higher than 0.5. The highest MCC score recorded is achieved with the RF model on bilirubin data, reaching a value of 0.74. An overview of the urine models' performances is presented in Table 4. Overall, the models trained on urine data show clearly less predictive power than the drainage models.

To improve explainability, the absolute values of the regression coefficients in the PLS-DA models were examined. For each fluid marker model, the values of each PLS coefficient were normalized to the sum of all of the 288 (number of wavelengths) × 3 (light pathways DT, AT, and AR) coefficients. Afterward, for each wavelength, the contributions of the coefficients for the different lightway paths (DT, AT, and AR) were summed to obtain a single value for each wavelength and fluid marker. In Figure 7A, the percentage contribution of the normalized sum of the PLS-DA coefficients is displayed in the form of a heatmap.

In Figure 7B, the RF feature importance values for each fluid marker model are examined. Similarly to the PLS-DA coefficients, the feature importances of the 3 lightway paths (DT, AT, and AR) were summed wavelength-wise. As the sum of all feature importances of the trained models is already equal to 1, there was no need to further normalize the contribution of each wavelength. The summed coefficients were then arranged into a heatmap.

A one-to-one comparison of the PLS-DA and RF heatmaps is not advisable because the PLS coefficients and the RF feature importances represent different statistical concepts (regression vs decision trees). Nevertheless, similar responses are visible when examining the RF feature importances and PLS-DA coefficients of the studied markers. For instance, the summed RF feature importances of drainage bilirubin exhibit 2 distinct peaks, located in the proximity of 470 and 570 nm. These peaks are also clearly visible for the RF model trained on urine bilirubin data. To some extent, similar responses can be seen in the PLS-DA coefficients of the same marker. Another noticeable example is given by hemoglobin. The bright peak at approximately 400 nm, visible in the summed RF feature importances of drainage hemoglobin, can also be identified in the PLS-DA coefficients of the marker. Interestingly, the same peak is clearly visible for the marker erythrocyte count on drainage data. Most probably, this is a consequence of the correlation between the number of blood cells in the samples and the total amount of hemoglobin [29,30].