| description abstract | Abstract. This study presents a data-driven framework for optimizing tympanostomy tube design to enhance fluid drainage efficiency and reduce drainage duration. Current commercial tubes often show inconsistent performance due to suboptimal fluid dynamics. The proposed approach integrates experimental data, neural network modeling, and computational optimization to refine key geometric and operational parameters. A total of 1080 experiments were conducted using 15 geometries, six materials (e.g., silicone, titanium), and two fluid types (water and Ciprodex). A neural network with two hidden layers (10 neurons each), trained via the Levenberg-Marquardt algorithm, predicted drainage outcomes of exiting droplet count and drainage duration with high accuracy (coefficient of determination (R2) = 0.80, root-mean-square error (RMSE) = 0.62 for droplet count; R2 = 0.92, RMSE = 2.19 for duration). Sensitivity analysis identified tube length, diameter, and inlet droplet count as the most influential parameters, while material and fluid type had limited impact. Simulated annealing (SA) (2800 iterations) was applied for multi-objective optimization, targeting maximal drainage and minimal time. Optimal designs featured increased length (1.27–1.4 mm), reduced diameter (<0.76 mm), and controlled inlet droplets (1 droplet). Experimental validation using three-dimensionally (3D)-printed PETG + PTFE prototypes confirmed model predictions, with drainage durations deviating by 7.5–10.5% and outlet droplet counts differing by at most one per trial. These results demonstrate the predictive accuracy and robustness of computational models, establishing a scalable methodology for integrating machine learning, global optimization, and low-cost prototyping in medical device design, supporting future in vivo studies and clinical translation. | |
