Medicament identity rather than total loading governs the morphology of electrospun poly(vinylpyrrolidone) nanofibers for regenerative endodontics: a machine learning analysis of a failure-inclusive dataset

Abstract

Electrospun fibers loaded with antibiotics or calcium hydroxide are being developed as intracanal carriers for regenerative endodontics, where the dose must stay low enough to spare the stem cells that repopulate the canal. Formulation development sweeps the medicament concentration while holding the polymer and machine settings fixed. We asked whether that sweep targets the right variable. We assembled ENDOSPIN-29, a dataset of 29 poly(vinylpyrrolidone) formulations produced under a single process backbone and loaded with metronidazole, ciprofloxacin, minocycline or calcium hydroxide, alone and in combination, retaining the five that produced no submicron fibers. Across nine regression models, those given per-medicament composition predicted fiber diameter far better than the same models given only total loading. The best reached a leave-one-out coefficient of determination of 0.84 and a median relative error of 18%, whereas every loading-only model performed at or below a mean baseline. Uniformity and distribution span behaved likewise; asymmetry was unpredictable. Dose response ran in opposite directions for different actives: ciprofloxacin thinned fibers monotonically from 406 to 257 nm, while metronidazole thickened them and destroyed fiber formation above 10% w/w. Holding out an entire medicament class removed the advantage, bounding the method to interpolation within a known drug panel.