[Feasibility study of a machine learning-based Raman spectroscopy model for pathological classification in secondary hyperparathyroidism].

Objective

To construct a classification model based on Raman spectroscopy (RS) and gradient boosting tree to differentiate the pathological types of diseased parathyroid glands in patients with secondary hyperparathyroidism (SHPT), and to evaluate its performance.

Methods

RS was collected from six types of normal cervical tissues (92 specimens in total) obtained from 5 cadaveric specimens, as well as from 82 parathyroid glands (787 spectra in total) resected intraoperatively from 23 patients with SHPT treated at Beijing Tongren Hospital, Capital Medical University(n=16), and Chaoyang Central Hospital, Liaoning Province(n=7). Using pathological diagnosis as the gold standard, a gradient boosting tree binary classification model was constructed to distinguish normal from diseased parathyroid glands, adenomatous hyperplasia from diffuse hyperplasia, and oxyphil cell-predominant from chief cell-predominant types. Leave-one-out cross-validation was employed, and model performance was evaluated using ROC-AUC, confusion matrices, and SHAP analysis, which also served to identify key Raman peaks.

Results

The model demonstrated excellent performance in distinguishing normal from diseased parathyroid glands (AUC=0.959). For different types of hyperplasia, the AUC for distinguishing adenomatous hyperplasia from diffuse hyperplasia was 0.920, with accuracies of 86.3% and 88.2%, respectively. For different cell types, the AUC for distinguishing oxyphil cell-predominant from chief cell-predominant types was 0.926, with a sensitivity of 92.9% and specificity of 76.3%. SHAP analysis revealed that characteristic peaks associated with proteins, lipids, nucleic acids, and carotenoids played key roles in the various classifications. Model stability analysis indicated good calibration, the validation-set ROC-AUC remained above 0.95 as the training sample size increased, and the log loss stabilized after approximately 60 iterations.

Conclusions

The combination of RS and the gradient boosting tree model can effectively differentiate the pathological types of diseased parathyroid glands in patients with SHPT, demonstrating high discriminative performance and good stability.