Automated Baseline-Correction and Signal-Detection Algorithms with Web-Based Implementation for Thermal Liquid Biopsy Data Analysis.

BACKGROUND/

objectivesDifferential scanning calorimetry (DSC) analysis of blood plasma, also known as thermal liquid biopsy (TLB), is a promising approach for disease detection and monitoring; however, its wider adoption in clinical settings has been hindered by labor-intensive data processing workflows, particularly baseline correction.

methodsWe developed and tested two automated algorithms to address critical bottlenecks in TLB analysis: (1) a baseline-correction algorithm utilizing rolling-variance analysis for endpoint detection, and (2) a signal-detection algorithm that applies auto-regressive integrated moving average (ARIMA)-based stationarity testing to determine whether a profile contains interpretable thermal features. Both algorithms are implemented in ThermogramForge, an open-source R Shiny web application providing an end-to-end workflow for data upload, processing, and report generation.

resultsThe baseline-correction algorithm demonstrated excellent performance on plasma TLB data (characterized by high heat capacity), matching the quality of rigorous manual processing. However, its performance was less robust for low signal biofluids, such as urine, where weak thermal transitions reduce the reliability of baseline estimation. To address this, a complementary signal-detection algorithm was developed to screen for TLB profiles with discernable thermal transitions prior to baseline correction, enabling users to exclude non-informative data. The signal-detection algorithm achieved near-perfect classification accuracy for TLB profiles with well-defined thermal transitions and maintained a low false-positive rate of 3.1% for true noise profiles, with expected lower performance for borderline cases. The interactive review interface in ThermogramForge further supports quality control and expert refinement.

conclusionsThe automated baseline-correction and signal-detection algorithms, together with their web-based implementation, substantially reduce analysis time while maintaining quality, supporting more efficient and reproducible TLB research.