Being minimally invasive, microneedles (MNs) are emerging as a promising technology for transporting active substances into the skin by bypassing the primary barrier, the stratum corneum. Comprising multiple MN arrays, MN array patches (MAP) penetrate the skin by forming microchannels, enabling direct diffusion of active compounds into the underlying layers. Consequently, the mechanical strength of MAP is a critical factor, as effective skin puncture is a prerequisite for successful drug delivery.

Mechanical strength is one of the cornerstones of MAP products to ensure that they remain structurally robust during application and product transport without breakage [1]. MAP is typically evaluated using a texture analyser, which applies a controlled compressive force onto the MAP. The common measured parameters include failure force (axial and transverse compression mode) and percentage reduction of MN length [[2], [3], [4], [5]]. In addition, the skin puncture performance is evaluated via membrane insertion tests to confirm the functionality of MAP products in effectively breaching the stratum corneum. Animal models, particularly porcine skin, have remained the mainstream for ex vivo/in vivo insertion and safety application evaluation of MAP products over recent decades [[6], [7], [8]]. Porcine ear skin is widely accepted as an industry-standard dermatological model due to its human-like skin structure and physiological responses [9,10]. However, biological variation within skin tissues remains one of the main obstacles in such evaluations [11]. To address this, an in vitro membrane model using Parafilm® was proposed and validated by Larrañeta, Moore, Vicente-Pérez, González-Vázquez, Lutton, Woolfson and Donnelly [4] to facilitate the prediction of ex vivo skin insertion. This model has been widely adopted as the first-line investigation for skin puncture performance [[12], [13], [14]] which allows the insertion efficiency and penetration depth to be evaluated based on the number of holes formed and layers penetrated for a quick preliminary assessment. Further histology and optical coherence tomography with ex vivo and in vivo skin insertion can verify the in vitro testing [15].

MAP fabrication is often a time-consuming process, particularly when using the conventional polymer casting method. Optimising the polymeric MAP formulation can be inefficient and costly, as it typically requires significant amounts of materials and active ingredients. Since the ability of MAP to effectively puncture the skin is critical, preliminary evaluations of MAP strength, both in vitro and ex vivo, serve as essential initial assessments of their functionality. Therefore, the development of a predictive model for MAP strength would be highly beneficial to accelerate and streamline the formulation design process.

Being a subset of artificial intelligence (AI), machine learning (ML) has recently received much attention in drug delivery for formulation design. By leveraging computer algorithms, ML enables automatic data analysis and pattern learning in order to generate new knowledge and meaningful outcomes. Unlike typical modelling, ML can handle vast and complex datasets to uncover meaningful patterns and make data-driven decisions. More importantly, ML can learn from past experiences to enhance its performance.

The capacity of ML to transform and expedite conventional formulation design and optimisation has been clearly established [16]. ML has been increasingly applied in various pharmaceutical formulation design and applications, including 3D printed formulations from hot melt extruded filaments [17], 3D printed orodispersible films [18], liposomes [19,20], nanoemulsions [21] and in vitro dissolution model [22]. MAP formulation development similarly reflects this trend, with growing interest in leveraging ML for enhanced design and performance [23].

Tarar, Aydın, Yetisen and Tasoglu [24] integrated finite element analysis (FEA) with ML to optimise key geometric parameters of MAP products, including length, inlet and outlet diameters, thickness and Bezier curvature, to enhance interstitial fluid extraction including human blood. The fluid behaviour was modelled using mechanical computational fluid dynamics simulations in COMSOL Multiphysics® and the optimisation was performed together using Bayesian optimisation algorithm in MATLAB to improve the retraction of maximum volumetric fluid. The group later refined the model by incorporating MAP material properties and skin-specific variables (e.g., age and gender), culminating in the development of a graphical user interface (GUI) for practical design application [25]. Building on this work, Abdullah and Tasoglu [26] subsequently focused on optimising MAP geometry to minimise pain, quantified through a computed pain index and visualised within a GUI. This model leveraged the LiveLink interface between COMSOL and MATLAB, combined with Bayesian optimisation, to solve complex optimisation problems.

Beyond geometric optimisation, Chumpu, Chu, Treeratanaphitak, Marukatat and Hsu [27] sought to expedite material selection for tapered-cone MAP designs by employing ML models trained on FEA data. By bypassing the need for time-intensive simulations, their approach enabled rapid prediction of mechanical performance, specifically the stress and strain distributions at the point of skin penetration based on data derived from 15 candidate MAP materials.

Using deep learning applied to digital image analysis, Rezapour Sarabi, Alseed, Karagoz and Tasoglu [28] developed a ML approach to detect defects in etched polylactic acid MNs fabricated by fused deposition modelling 3D printing. Additionally, a GUI was created to predict MAP quality based on geometric parameters (base diameter, height and drafting angle) and etching variables (potassium hydroxide solution concentration and etching duration).

Apart from MAP design and formulation, several studies have employed ML for predicting drug permeation from MAP. Yuan, Han, Yap, Kochhar, Li, Xiang and Kang [29] combined mechanistic models (Fick's law of diffusion) and ML models to predict the permeation amount and percentage of different classes of permeants (metal ions, small and big molecules). Several features including skin type (rat and human), MAP type (hydrogel and plastic), MN length, MAP surface area, drug loading, drug permeation time and drug molecular weight (MW) were considered in the prediction. Using the same data set, Biswas, Dhondale, Singh, Agrawal, Muthudoss, Mishra and Kumar [30] improvised the ML models by using voting regression model and developed a GUI for the prediction.

Nevertheless, a critical gap remains in current ML applications for MAP formulations, particularly in linking mechanical performance to formulation design. This is especially important for polymeric MAPs, owing to their rising prominence, user-friendly application and versatile capacity to incorporate active compounds [[31], [32], [33], [34]]. Building on the advantages of ML, the present study aims to predict the evaluation of polymeric MAP strength — specifically, failure force and skin insertion profiles — using formulation data gathered from existing literature. To achieve this, the study focuses exclusively on polymeric MAP products loaded with or without active substances in their native form. Reports involving actives incorporated as nanoparticles, liposomes, nanoemulsions or other forms were excluded as such formulations introduce substantial variability and an additional layer of complexity to the ML framework. With this, the study centres on key determinants known to significantly influence MAP strength, including the types of polymers, active substances and MN dimensions.