<b>Machine Learning Algorithms Interpreting Circulating MicroRNA Signatures for Real-Time Monitoring of Chemotherapy Response: A Randomized Controlled Trial</b>

Abstract

Background: The standard waiting period of six to eight weeks for radiographic assessment of chemotherapy response creates a critical clinical gap during which non-responsive tumors may progress and responsive patients endure unnecessary toxicity. Circulating microRNAs offer dynamic biological readouts of treatment effect, but their complex, high-dimensional patterns have eluded conventional statistical interpretation. Objective: To determine whether a machine learning algorithm analyzing serial circulating microRNA levels can predict radiographic chemotherapy response at two weeks with acceptable concordance compared to standard week-eight imaging. Methods: This parallel-group randomized controlled trial enrolled 110 adults with advanced solid tumors receiving first-line platinum-based or taxane-based chemotherapy. Participants were randomized to an intervention group (serial microRNA profiling with algorithmic analysis at baseline, day 3, day 7, and day 14) or a control group (standard care with imaging at week eight; blood samples stored for post-hoc analysis). The primary outcome was concordance between the algorithm's two-week prediction and week-eight RECIST-defined response, assessed using positive percent agreement and negative percent agreement with 95% confidence intervals. Secondary outcomes included time to treatment switch and six-month progression-free survival. Results: Among 101 participants completing the trial, the algorithm demonstrated 85.0% overall concordance (95% CI 76.5%–91.4%), with positive percent agreement of 88.0% (95% CI 75.7%–95.5%) and negative percent agreement of 82.0% (95% CI 68.6%–91.4%). One participant with indeterminate algorithm output was excluded from concordance calculations. A significant time-by-group interaction was observed (F(3, 297)=14.82, p<0.001, with Greenhouse-Geisser correction). Time to treatment switch was shorter in the intervention group (median 23 days, IQR 20–28 vs. median 42 days, IQR 36–48; log-rank p<0.001), and six-month progression-free survival was higher (76.0% vs. 58.8%, p=0.04). Conclusion: Machine learning analysis of serial circulating microRNA signatures predicted chemotherapy response at two weeks with high concordance compared to standard imaging, enabling earlier treatment switches and improved progression-free survival. These findings provide proof of concept for real-time microRNA-based monitoring, though confirmation in a multicenter trial with algorithm-guided treatment decisions is required before clinical implementation.

Ähnliche Arbeiten