Evaluation of multiple machine learning models predicting the results of hybrid imaging in primary
Anna Drynda1, Jacek Podlewski2, Karolina Kucharczyk3
1Doctoral School of Medical Sciences and Health Sciences, Jagiellonian University Medical College, Kraków, Poland.
Background:
Primary hyperparathyroidism (PHP) diagnosis is based on abnormalities in biochemical blood tests. Preoperative localization of the affected gland with imaging may increase the effectiveness of the surgical treatment. The aim of this study is to evaluate predictive strategies for the assessment of radiotracer uptake in pre-operative [99mTc]Tc-sestamibi scintigraphy ([99mTc] Tc-MIBI SPECT-CT) among PHP patients to identify individuals with a high probability of negative results, and to develop clinical decision-making tools.
Material And Methods:
Development and evaluation of logistic regression (LR), classification trees utilizing the classification and regression trees (CART) algorithm, random forest (RF), and boosted trees employing XGBoost (XGB) predictive models. All models were constructed using data obtained from 499 patients diagnosed with PHP who underwent [99mTc]Tc-MIBI SPECT-CT imaging between 2010 and 2022 at the University Hospital in Cracow, Poland.
Results:
The LR model demonstrated the best out-of-sample performance, achieving a specificity of 81.3% and an accuracy of 69.3%, with a sensitivity of 55.7%. Along with CART and XGB, LR performed well when using only 5 predictors: concentrations of parathormone (PTH), serum calcium, serum phosphates, total serum vitamin D, and maximal lesion diameter measured in ultrasound. Random forest (RF) exhibited higher sensitivity (62.7%), but lower specificity (74.2%) and accuracy (68.6%). Other models demonstrated subpar performance.
Conclusions:
Logistic regression and RF models were the most effective in predicting radiotracer uptake in pre-operative hybrid imaging of the parathyroids, suggesting their suitability as the foundation for software to be used in clinical settings. However, opting for the CART model, despite its easier interpretation, would come at the expense of performance.
