Machine learning integration of thermal liquid biopsy and clinical biomarkers for pancreatic cancer diagnosis and prognosis: A retrospective cohort study

This is the authors' abstract. We don't add key points for this paper.

Background Pancreatic cancer continues to be one of the most challenging malignancies to diagnose and treat, and it is often detected at an advanced stage owing to asymptomatic early phases and limitations in current diagnostic tools, highlighting the need for improved diagnostic methodologies. Methods We aimed to evaluate the potential for integrating information from thermal liquid biopsy (TLB) thermograms with clinical biomarkers to improve the diagnosis and prognosis of pancreatic cancer. Serum samples from 381 Danish patients with pancreatic cancer and 325 patients referred with non-organ specific signs and symptoms of cancer, but without confirmed cancer after clinical and computed tomography evaluation, were analyzed. TLB thermograms provided information on the partial excess heat capacity of the serum as a function of temperature. Three classification models were constructed using machine learning algorithms based on variable selection with penalization, applying cross-validation and resampling techniques: iClin Model (age, Eastern Cooperative Oncology Group Performance Status, carbohydrate antigen 19.9, and C-reactive protein), iTLB Model (discordant pairs of temperatures from thermograms), and iTLB + iClin Model (discordant pairs of temperatures and clinical data). Results The iClin Model achieved high diagnostic performance, with a validation area under the curve (AUC) of 0.95 ± 0.01 for differentiating symptomatic controls from patients with pancreatic ductal adenocarcinoma (PDAC), but showed no significant association with overall survival (OS). The iTLB Model showed limited diagnostic performance, with a validation AUC of 0.59 ± 0.03, but was associated with OS in patients with PDAC. The combined iTLB + iClin Model preserved high diagnostic performance, with a validation AUC of 0.95 ± 0.02, sensitivity of 96.20%, specificity of 91.20%, positive predictive value of 90.48%, and negative predictive value of 96.59%. For early-stage PDAC, the iTLB + iClin Model achieved an AUC of 0.93 (95% CI: 0.89–0.98), compared with 0.96 (95% CI: 0.95–0.98) for stages III–IV. Conclusion This study demonstrates the potential of combining specific clinical biomarker information with advanced techniques such as TLB, to improve the accuracy of pancreatic cancer diagnosis and prognosis.