Accessible Digital Reconstruction and Mechanical Prediction of 3D-Printed Prosthetics

Customized prosthetics can significantly improve the lives of patients suffering from amputation. However, access to customized prosthetics is often limited in resource-constrained communities [1,2]. One significant aspect of prosthetics aggravating this limitation is the need for customization. For prosthetics to be effective, they must be customized to each patient’s unique anatomy and functional needs [3–7]. Demand for customization is especially highlighted by young patients undergoing physical development and early-stage patients whose residual limb shape is changing [8–10]. This demand for unique products customized for individual patients reduces accessibility. This limitation can be exacerbated in resource-constrained communities. According to the World Health Organization (WHO), only one in ten people in need can access assistive products, including prostheses, due to their high cost, lack of awareness, unavailability of trained personnel, policy, and financing issues [11]. An analog method is widely adopted to produce customized prosthetics in resource-constrained communities. In this method, professionals manually shape thermoforming materials to fit the physical model of the residual limb cast from the plaster mold [12]. Although it is widely used, it faces several limitations. First, patients must visit the hospital multiple times to collect data on the residual limb and refine their model because the patient’s presence is required in the casting and adjusting of the model when using plaster. Second, the quality of prosthetics heavily relies on practitioners’ level of skills. Therefore, digital technologies and additive manufacturing have been introduced to overcome these limitations. Digital technologies and additive manufacturing can reduce the time and effort required for designing and producing customized prosthetics while increasing convenience and quality [12]. However, the digital approach has not been widely implemented in resource-constrained communities for multiple reasons. First, the digital reconstruction process is often limited in resource-constrained communities. Many researchers proposed a digital reconstruction method of the residual limb [7,13,14]. However, they require medical equipment such as magnetic resonance imaging (MRI) and 3D scanners, which are inaccessible in some communities and require extensive training for their use [15–17]. Therefore, we propose a method to achieve a high-quality digital model of the residual limb from smartphone videos. Smartphones are user-friendly and accessible devices in resource-constrained communities [18]. Therefore, utilizing smartphones to capture the residual limb’s geometry can reduce the equipment and educational barrier in digital reconstruction [15,16,19]. Second, determining the prosthetic design, filament material, and printing process is difficult because predicting the mechanical functionalities of the 3D-printed prosthetic is challenging. Especially, research on the influence of printing process on the mechanical functionalities of 3D-printed prosthetics has been minimal despite the wide use of 3D-printing technology in prosthetic production [12,15,19–21]. This limitation can be critical in resource-constrained communities with limited access to high-end 3D-printing technology and filaments such as metal or carbon 3D printing. In this research, we focus on fused deposition modeling (FDM), also known as fused filament fabrication (FFF), because it is known to be one of the most accessible and economical additive manufacturing methods [7,8]. FDM faces a challenge that the filament material and printing process significantly influence the mechanical properties of 3D-printed parts [22–26]. Many researchers aim to predict the mechanical properties of 3D-printed parts via finite element analysis (FEA) to overcome this challenge [27,28]. However, FEA-based methods require a unique computer-aided design (CAD) model for each set of printing process parameters, making exploration of the printing process inefficient. To overcome this challenge, we generate a dataset and implement a machine learning model XGBoost to predict the effective material data of 3D-printed parts, which can be used to predict the mechanical functionalities of the 3D-printed prosthetic. Overall, we aim to increase the accessibility of 3D-printed prosthetics by (1) obtaining a digital model of the residual limb from a smartphone and (2) implementing data-driven methods for efficient prediction of the mechanical functionalities of 3D-printed prosthetics. The method promotes using smartphones and 3D-printing technologies, particularly FDM, to enhance accessibility in prosthetic design and production.

Digital reconstruction is a method of reconstructing a digital model from real-world images or videos. Digital reconstruction of the residual limb is essential in customized prosthetic design using CAD because customized prosthetics are designed according to the digital residual limb model. Therefore, many researchers are vying to capture the residual limb’s details accurately [7,13,14]. Rai et al. [14] used MRI to capture details of the residual limb. Górski et al. [13] used a 3D scanner to reconstruct the digital model. Although they achieved an accurate digital residual limb model, they used medical equipment such as MRI and 3D scanners, which are sometimes inaccessible in resource-constrained communities and require intense education [15–17]. This limitation discourages the adoption of these methods. In contrast, smartphones can be an accessible device for digital reconstruction. According to surveys, the number of mobile Internet users reached 320 million people, 27% of the population in sub-Saharan Africa in 2023, and it is expected to rise to 320 million people, 37% of the population in 2030 [18]. Therefore, many researchers adopted smartphones and photogrammetry to achieve digital representations of the residual limb. Cullen et al. [17] utilized a genetic algorithm to reconstruct the residual limb from smartphone images and photogrammetry. Lee et al. [15] used smartphone images and photogrammetry to reconstruct the residual limb and adopted a parametric design to generate prosthetics from the reconstructed model. Their approach relied on photogrammetry for reconstruction, which requires manual parameter tuning for each case and expertise in digital reconstruction. Also, they use high-end smartphones to capture images of the residual limb, which increases barriers to digital reconstruction. In contrast, the proposed method eliminates the need for parameter tuning and demonstrates its feasibility on resource-constrained communities through a case study conducted with accessible smartphones.

Many researchers have introduced methods to reconstruct digital models from images and videos. Zhang et al. [29] proposed NeRS, and Wang et al. [30] proposed Nerf to reconstruct digital models from multiple views. However, their method could be improved in terms of mesh quality. Wang et al. [31] proposed Neus to achieve a higher-quality surface, but their method requires camera path information. Camera path information is often inaccessible with the accessible smartphones in resource-constrained communities. Finally, Li et al. [32] proposed Neuralangelo to recover detailed structures of real-world scenes from a video and images. Nueralangelo can achieve high-quality mesh without parameter tuning for different cases, which reduces the manual effort required in the digital reconstruction process. Therefore, the 3D reconstruction method in this research implements Neuralangelo to reconstruct the patient’s residual limb from a smartphone video.

Many factors influence the reconstructed mesh quality. One primary factor is the video quality, which is strongly affected by the type of smartphone. Ganeeva and Myasnikov [33] investigated widely used 3D reconstruction methods and identified the influence of input video quality on reconstructed mesh quality. Multiple factors can influence the video quality of smartphones. The study by Joskowicz and Ardao [34] provides insights into the intricate influences of frame rate, bit rate, display size, and video content on video quality. Uhrina et al. [35] investigated the effects of compression methods on video quality. Unfortunately, access to high-end smartphones is not always guaranteed in resource-constrained communities. Therefore, we should validate the reconstruction method using accessible smartphones within the community to confirm its feasibility. Another primary factor is lighting. Kemelmacher-Shlizerman et al. [36] highlighted the implications of lighting variations on facial geometry reconstruction. Gómez-Gutiérrez et al. [37] compared photo reconstruction methods and found that shadows and lighting variations critically affect the density and accuracy of point clouds produced from photogrammetry. Belhaoua et al. [38] discovered that intensity, directionality, and light uniformity across the scene all collectively influence the quality of captured images and subsequent reconstruction accuracy. Therefore, we must validate the method on various lighting conditions to ensure feasibility.

FDM, also known as FFF, is a technology that realizes an object by depositing layers of thermo-plastic extrusion supplied from a filament [5,6]. In this research, we focus on FDM because it is known to be one of the most accessible and economical additive manufacturing methods [7,8]. FDM has been widely adopted in the medical and dental fields to rapidly and economically produce customized medical products such as prosthetics, orthotics, and artificial organs [39–41]. For example, Fenollosa-Artés [42] 3D-printed bones with multiple materials using FDM. Brancewicz-Steinmetz et al. [43] employed FDM to 3D print upper limb prosthetics and proposed a method to improve the prosthetic’s surface quality. Lestari et al. [44] used the Taguchi method and the response surface to optimize the 3D-printing process for the prosthetic socket design. However, their work has limitations. First, they do not identify the influence of the 3D-printing process on the mechanical functionalities. In FDM, the printing process parameters influence printed parts’ mechanical properties [22–26]. Second, their work relies on analyzing physical 3D-printed prosthetics. Although this approach has the advantage of capturing data that closely represents 3D-printed prosthetics, it may limit the dataset’s size, range, and continuity due to the challenges of expanding it. Therefore, a data-driven method encompassing a wide range of filament material and printing process parameters is needed to determine the influence of printing process on mechanical functionalities.

Printing process parameters such as infill pattern, infill density, layer height, line width, wall thicknesses, and building direction are known to influence the mechanical properties of 3D-printed parts [22–26]. Therefore, the influence on the mechanical functionalities of 3D-printed prosthetics should be considered when selecting the printing process parameters. This is especially critical in lower-limb prosthetics, which need to withstand the weight and movement of patients throughout their daily use. However, research on optimizing the 3D-printing process has been minimal despite its importance [12,20]. One major challenge in analyzing 3D-printed prosthetics is its micrometer details of deposited filaments, inevitable when 3D printing. Preparing computing resources for FEA on a CAD model with excessive micrometer details is difficult, especially in resource-constrained communities. One popular method to significantly reduce the details of the CAD model by approximating the mechanical properties of repetitive structures is homogenization [27,28,45]. Homogenization achieves effective material data, which refers to approximated mechanical properties of repetitive structures in a format of anisotropic material data, by conducting FEA on repetitive volumetric elements (RVEs), a portion of repetitive geometry that can represent the whole geometry by repeating itself. However, one limitation of this method is that it requires a different CAD model of RVEs for different printing processes. Therefore, we generate effective material data from various filament material data and printing process parameters via homogenization. The machine learning model trained on the generated data can predict the effective material data from filament material data and the printing process without the CAD model of RVEs.

We propose methods to increase the accessibility of 3D-printed customized prosthetics. First, we propose a method for reconstructing a digital model of the residual limb from smartphone videos that does not require parameter tuning for different cases. Clinicians can capture videos of residual limbs in primary care clinics with smartphones. Clinicians can also obtain supplementary information about the patient, such as dimensions of the residual limb, weight, medical status, and prosthetic functionality requirements. Afterward, clinicians can send the video and supplementary information to central computational resources to reconstruct the residual limb and generate the prosthetic design. Central computational resources are necessary because the computational resources required for digital reconstruction exceed what an accessible smartphone can handle. Second, we propose a method to predict the mechanical functionalities of 3D-printed prosthetics by implementing a machine-learning model trained from the generated dataset. Prosthetic designers can use the prediction results, along with user requirements and the 3D-printing capabilities of fabrication centers, to determine design parameters, filament materials, and printing process settings. The final 3D-printed prosthetic is delivered back to the primary care for fitting and tracking. Importantly, all procedures between capturing the residual limb video and fitting the prosthetic can be performed remotely, significantly increasing accessibility. The overall process for achieving a 3D-printed prosthetic is presented in Fig. 1. The following subsections describe the method and evaluation process for digital reconstruction and mechanical functionality prediction of the 3D-printed prosthetic.

The digital model of the residual limb is reconstructed from a video captured with a smartphone, which is a user-friendly and highly accessible device [18]. Utilizing smartphones to capture the residual limb’s geometry reduces the equipment and educational barriers in digital reconstruction [15,16,19].

A good-quality video is essential for obtaining a high-quality residual limb mesh while reducing the required time and computational resources. The official tutorial for Colmap, a key component of Neuralangelo, a primary element of the proposed reconstruction pipeline, provides the following guidelines [46,47]. First, capture images with distinct textures and avoid entirely texture-less areas. Second, ensure images are captured under consistent lighting conditions. Third, capture images with significant visual overlap to ensure seamless reconstruction. Last, capture images from varied viewpoints for comprehensive coverage. On top of that, we propose three more guidelines based on experience. First, the smartphone should rotate one lap around the residual limb. Second, the length of the video should be around 20–30 s. Longer videos with multiple laps can also be used for reconstruction, but it will likely increase the reconstruction time and computing resources without improving mesh quality. Rotating the smartphone faster than the guideline around the object may result in poor-quality meshes due to video blurring. Third, the video must always include both the top and side views of the residual limb. If the video fails to capture specific parts of the residual limb, those parts might not be well reconstructed. Figure 2 illustrates the smartphone’s recommended path for capturing a video of the residual limb.

The reconstruction process involves three main steps. First, images are sampled from the video at three-frame intervals, and backgrounds are removed using the RemBG package [48]. Removing the background ensures that the 3D mesh focuses solely on the residual limb, excluding background elements from the reconstructed scene. Second, we reconstruct the mesh from processed images using Neuralangelo [32]. Third, the reconstructed mesh goes through postprocessing. Postprocessing eliminates residual artifacts from the background and shadows, reduces surface noise, and repositions the mesh to a centered, upright orientation [15]. The postprocessing process is scripted in CAD programs to reduce manual efforts. Iterative Laplacian smoothing is applied to reduce mesh noise. The overview of the digital reconstruction process is illustrated in Fig. 3.

The accuracy of the reconstructed mesh is assessed by calculating the average deviation from the original mesh [49,50]. The average deviation is the average distance between the original mesh’s points and the reconstructed mesh’s closest points. Points are randomly generated on each mesh, with an equal number for both, and the distance is measured. The number of points is increased until the average deviation converges. This evaluation method benchmarks Chamfer-L1 distance, one of the standards for evaluating similarities between two sets of points by calculating the sum of the nearest distances between points in each set [32,51,52]. In contrast to the Chamfer-L1 distance, where the nearest distances are measured between points on the point cloud, the nearest distances are measured between points randomly generated on the mesh in average deviation.

Parametric design allows designers to adjust designs by altering design parameters instead of redoing the 3D modeling process and provides design space represented in numerical values for analysis and optimization [53]. Therefore, parametric design is widely used to generate various prosthetic designs from the digital model of the residual limb [15,21]. However, one of the challenges in parametric prosthetic design is ensuring the mechanical functionalities of the generated design. This challenge is particularly critical for 3D-printed products, as accounting for the influence of the 3D-printing process on mechanical functionalities is challenging. To overcome this challenge, we (1) generate a dataset of effective material data from various combinations of filament materials and printing process parameters and (2) train an XGBoost model that predicts the effective material data from filament material data and printing process parameters. The predicted effective material data can be applied to a CAD model of the prosthetic for FEA to predict the mechanical functionalities. The prosthetic design, filament material, and printing process parameters can be adjusted if the predicted mechanical functionalities do not meet the requirements. The trained XGBoost model can be reused for different patients if the filament material and printing process parameters are within the generated dataset.

A dataset capturing the relationship between filament material data, the printing process parameters, and the effective material data is generated. Data generation follows three key steps. First, generate data points representing various combinations of filament material data and printing process parameters. Table 1 presents the variable ranges of the dataset. Young’s modulus, tensile yield strength, and Poisson’s ratio are selected for material data, and layer height is selected for the printing process parameters. Layer height is known to influence the mechanical properties in the vertical direction [23,26]. Other printing process parameters, such as infill pattern, density, and wall thicknesses, may also affect mechanical properties [54]. However, this research does not consider them because the prosthetic socket in this article is designed as shells that encase the residual limb. The range of material data is selected to cover most of the known polymers [55]. A total of 10,000 data points are generated using the Latin hypercube sampling (LHS) method to sample the data uniformly. Second, generate the RVE CAD models from the printing process parameters via parametric design. The cross section of the printed filament is adopted from the study by Coogan and Kazmer [56], where layer height and line width determine the cross-sectional dimensions. The line width of the extruded filament is fixed at 0.4 mm because the standard diameter of the filament extruder of the FDM 3D printer is 0.4 mm. Third, derive effective material data from FEA results and geometry of the RVE following the definition of each effective material data. Effective material data consist of eight values: Young’s modulus, shear modulus, tensile yield strength, and Poisson’s ratio in both horizontal and vertical directions. The mechanical properties of extruded filament are anisotropic because the reaction force varies depending on the direction of the applied force. The overall process for data generation is illustrated in Fig. 4.

XGBoost is selected as a machine-learning model for prediction because it is one of the most effective models for predicting the mechanical properties of 3D-printed parts, based on its printing process [57]. The XGBoost model takes Young’s modulus, tensile yield strength, Poisson’s ratio of the material, and layer height as input and predicts effective material data: Young’s modulus, shear modulus, tensile yield strength, and Poisson’s ratio in the horizontal and vertical directions. The XGBoost is trained with the following parameters: $eta=0.05$⁠, max $depth=5$⁠, estimator $number=1000$⁠. Root-mean-square error (RMSE) and $R2$⁠, widely used metrics for evaluating regression tasks, are employed as performance evaluation metrics [58]. The $k$-fold cross-validation method (⁠$k=5$⁠) is adopted to evaluate the trained XGBoost model [59].

The effective material data predicted from the filament material and the printing process are applied to the prosthetic’s CAD model when conducting FEA to evaluate the mechanical functionalities of prosthetics printed according to the given condition. The mechanical functionality of the prosthetic is defined by the ultimate strength in ISO-10328 P4 loading condition 2, which represents the instant of maximum loading occurring late in the stance phase of walking based on the locomotion data of patients under 80 kg [60,61]. Although ISO-10328 is a physical test standard, FEA is often used to simulate the evaluation in digital [61,62]. Failure of the prosthetic is determined based on the lowest tensile yield strength in any direction. Employing FEA for evaluation is less accurate than the physical test. However, its test speed and cost advantages enable the exploration of various filament materials, designs, and printing processes with reduced time and resources. Predicting the mechanical functionalities of 3D-printed prosthetics can help professionals and patients determine filament materials, designs, and printing processes.

The case study demonstrates the method’s feasibility. It represents designing a customized prosthetic socket for a patient with below-knee amputation on the right leg using an accessible smartphone in Rwanda and Kenya and an FDM 3D printer with polymer filaments. The 3D-printed generic residual leg is used as a substitution for the patient’s body, and no actual patient data were collected or used in the study. Figure 5 illustrates the digital model and 3D-printed generic residual limb used for the case study and the 3D-printed prosthetic socket generated from the method. Original Prusa XL is used to 3D print the generic residual limb and the generated socket. ASA natural filament from Prusament is used to 3D print the generic residual limb, and Matte PLA Chocolate from Ziro is used to 3D print the prosthetic socket.

Case studies of digital reconstruction are conducted in eight different conditions following the full-factorial design of the experiment. Eight different conditions are combinations of three two-level factors. The first factor is the type of smartphone. Ikosora+ from MTN and Neon Ray from Safaricom are selected to represent accessible smartphone devices in the community. Ikosora+ is a smartphone from local vendor MTN in Rwanda, which the government sponsors to increase smartphone accessibility and availability [63,64]. Neon Ray is one of the most accessible smartphones from Safaricom, a local vendor in Kenya. The video quality of Neon Ray is 720p, and the video quality of Ikosora+ is not provided. The second factor is lighting. Each smartphone took a video of the 3D-printed generic residual limb in two different lighting conditions, each representing a dark and bright indoor environment. Bright environments are light over 300 lux, and dark environments are under 300 lux. According to EN-12464, 300 lux is the minimum light required for moderate industrial work. The third factor is the distance between the person filming the video and the residual limb. We draw a circle around the residual limb and follow the circle to film a video of the residual limb. The radius of the circle is either 45 cm or 65 cm. All meshes are reconstructed using NVIDIA GeForce RTX 4090 with 24GB VRAM.

Table 2 presents reconstruction time, scale, and average deviation for each reconstruction. Reconstructed results are presented in Fig. 6. The results show that different conditions result in different mesh scales and qualities. Seven out of eight results showed an average deviation of less than 2 mm between the original and the reconstructed mesh. According to a previous study, a mesh with an average distance of less than 2 mm is viable for designing prosthetics [49]. Only condition (h), filmed with Ikosora+ from a long distance and a dim light, shows an average deviation of 2.1 mm. The best-quality meshes achieved an average deviation of 1.1 mm. These meshes are reconstructed from a video filmed with a Neon Ray and a closer distance. The average mesh construction time is 21.6 h, with the fastest time of 20.2 h and the slowest time of 26.9 h. The reconstruction scale varies, but it remains consistent for the same type of phone and distance. The value of the reconstruction scale may also differ depending on the CAD tool and mesh processing pipeline.

Table 3 presents the Spearman correlation coefficient and its $P$-value between the video and the reconstruction results. The conditions for capturing videos are encoded to categorical values in the analysis. The correlation coefficient of the type of smartphone and the average deviation with the original mesh is 0.88, with a $P$-value of 0.003. This indicates that the type of smartphone significantly influences the mesh quality. The correlation coefficient of distance and reconstruction scale is $−0.90$⁠, with a $P$-value of 0.002. This indicates that the scale of the reconstructed mesh increases as the smartphone gets closer to the residual limb. The correlation coefficient of the type of smartphone and the video length is 0.89, with a $P$-value of 0.003. This indicates that the type of smartphone influences the reconstruction time. One probable explanation for this result is that different smartphones have different frames per second, which causes a difference in the number of images sampled from the video. This explains that smartphones with high frames per second can result in a longer reconstruction time. Other influences of video conditions on reconstruction results are trivial. These results show that selecting the right smartphone type and capturing distance are crucial for achieving a high-quality digital representation.

This section demonstrates predicting the mechanical functionality of 3D-printed sockets to determine their design, filament material, and printing process. It uses a prosthetic socket design generated from the digital model of the generic residual limb reconstructed in the previous section.

The $R2$ and RMSE of XGBoost validated in the $k$-fold cross-validation method (⁠$k=5$⁠) are presented in Table 4. All predictions achieved $R2$ over 99% with low RMSE. The trained XGBoost model can accurately predict the effective material data of 3D-printed parts when the filament material data and printing process parameters are given.

The mechanical functionalities of different prosthetic designs printed in different filament materials and printing processes are evaluated following FEA simulation of ultimate strength in ISO-10328 P4 loading condition 2, which represents the instant of maximum loading occurring late in the stance phase of walking based on the locomotion data of patients under 80 kg [60,61]. The wall thickness of the prosthetic socket is controlled as a design parameter. The effective material data of parts 3D printed from different filament materials and printing processes are predicted using the XGBoost model trained in Sec. 4.2.1 and applied to different designs of prosthetic socket’s CAD models. Figure 7 shows an example of an FEA result. As depicted in the figure, the prosthetic socket fails in the bottom part when force is applied following ISO-10328 loading condition 2.

One hundred combinations of socket design, filament material, and printing process are generated via the LHS method from variable ranges in Table 5, identical to Table 1 with one more row of design parameters: wall thickness of the prosthetic socket. The predicted safety factor of each combination is presented in Fig. 8 according to the tensile yield strength of the filament material, layer height, and wall thickness. The results of the analysis show that stronger filament material provides stronger prosthetics. However, the stronger filament may be more expensive. A prosthetic becomes stronger with thicker walls. However, thicker walls increase the prosthetic’s weight and printing time. Small layer height strengthens the prosthetic because it makes the 3D-printed part stronger in its vertical direction [23,26]. However, this will increase the time needed to print the socket [21]. Patients, clinicians, and designers can determine the prosthetic design, filament material, and printing process of 3D-printed prosthetics according to predicted mechanical functionalities, the patient’s weight, medical conditions, and lifestyle.

We propose methods to increase the accessibility of 3D-printed customized prosthetics. First, we propose an accessible method for the digital reconstruction of the residual limb. We specifically (1) provide guidelines for capturing the video of the residual limb with smartphones and (2) augment Nueralangelo for digital reconstruction of the residual limb. Second, we propose a method to predict the mechanical functionalities of the customized prosthetic printed in the FDM 3D-printing method, aiding decisions on filament material, prosthetic design, and printing process. We (1) generate data on effective material, considering the filament material data and printing process parameters, and (2) train the XGBoost model to predict the effective material data using the generated data. The predicted effective material data can then be integrated into the prosthetic’s CAD model, generated from the reconstructed digital model of the residual limb to predict the mechanical functionalities of the 3D-printed prosthetic.

We also present a case study to show the feasibility of the method. The case study illustrates a scenario where a customized prosthetic socket for a patient with below-knee amputation on the right leg is designed using an accessible smartphone and an FDM 3D printer with polymer filaments. The case study consists of two major parts. The first part of the case study focuses on the digital reconstruction. The results show that seven out of eight results have an average deviation from the original model of less than 2 mm, which indicates that mesh is viable for designing prosthetics [49]. The best-quality mesh showed an average deviation of 1.1 mm. The Spearman correlation coefficients between the videos and the reconstructed results show that the type of smartphone used to capture the video influences the mesh quality and reconstruction time, with a correlation coefficient of 0.89 (⁠$P$-value $=0.003$⁠) and 0.88 (⁠$P$-value $=0.003$⁠), respectively. One probable explanation for the smartphone type’s influence on reconstruction time is that different smartphones have different frames per second, leading to a difference in the number of images sampled from the video. According to this explanation, smartphones with higher frames per second can improve the mesh quality but result in a longer reconstruction time. The distance between the residual limb and the smartphone camera also showed a high correlation with a coefficient of $−0.90$ (⁠$P$-value of 0.002). This indicates that the reconstructed mesh gets bigger as the smartphone moves closer to the residual limb. The case study on the digital reconstruction has two major implications. First, it reduces the manual effort of professionals in the reconstruction. Previous studies employing photogrammetry for reconstruction require experts to manually set the bounding box and select parameters for different reconstruction cases. In contrast, the proposed reconstruction method eliminates the need for such efforts. Second, the analysis results of the digital reconstruction show the influence of various conditions, such as smartphone type, lighting conditions, and capturing distance, on the mesh scale, quality, and reconstruction time. These results indicate that selecting the right smartphone type and capturing distance is crucial for achieving a high-quality digital representation. The second part of the case study covers designing a prosthetic socket and selecting the filament and printing process parameters. It demonstrates the considerations for selecting these factors based on the mechanical functionality predictions of the 3D-printed sockets. The XGBoost model trained on the generated dataset can accurately predict effective material data using filament material data and printing process parameters, achieving an $R2$ value exceeding 0.99 for all effective material data. The predicted effective material data can be assigned to the CAD model of prosthetics for FEA to predict its mechanical functionalities. Mechanical functionality predictions of 100 different 3D-printed prosthetic sockets show that different combinations of socket design, filament material, and printing process can result in different mechanical functionalities. Therefore, patients, clinicians, and prosthetic designers can determine these factors in 3D-printed prosthetics according to predicted mechanical functionalities, the patient’s weight, medical conditions, and lifestyle.

However, the presented method has several limitations. First, it assumes that material data are provided, but only specimen test results are provided in the commercial filament’s datasheet [54]. Specimen test results differ from the original filament material data because the printing process of the specimen influences the mechanical properties of the 3D-printed parts. Therefore, the printing process of the specimen must be considered when using the specimen test results as a description of filament material. Second, we adopt homogenization for prediction. While homogenization significantly reduces the details of the CAD model, reducing the required time and computational resources for FEA, it approximates the mechanical properties of the repetitive structures. This approximation may introduce inaccuracies. Third, the dataset is generated via FEA. Although FEA is widely used for mechanical properties prediction of 3D-printed parts, it cannot simulate the imperfect bonds between filaments [23]. Therefore, the 3D-printed prosthetics may be weaker than the predicted strength. One way to mitigate this issue is to introduce a new parameter that explains the bonding perfection between filaments. Third, this research is limited to presenting case studies under specific situations. Further studies are needed to evaluate generalizability across a wider range of residual limb shapes, smartphone types, lighting conditions, and capturing distances. Fourth, the presented method does not account for the complexities of implementation at different scales. Factors such as the demand for computational resources, availability of 3D printers, and the need for skilled professionals can vary significantly depending on the community’s population, demand for prosthetics, and existing medical infrastructure. Future research should explore optimizing these resources to adapt the method effectively to diverse settings.

The case study also has some limitations. First, the residual limbs’ ends and scales are manually identified, which limits the seamless integration of digital reconstruction and design generation processes. Second, the FEA used to predict the mechanical functionality of the prosthetic is performed manually. 3D-printed prosthetics have unique shapes customized for individual patients, which changes the geometry and boundary condition selection for FEA. Therefore, FEA is manually done to consider each prosthetic’s unique shape. This limits the efficient exploration of prosthetic designs, filament materials, and printing processes. One future research direction could be to propose a multimodal AI prediction model that uses the geometry of the patient’s residual limb or customized prosthetic as input to predict the mechanical functionality of prosthetics in various shapes. Third, the prosthetic design and printing process are explored within a limited range. Prosthetic wall thicknesses of 2 mm, 4 mm, and 8 mm are selected as design parameters. Layer heights of 0.1 mm, 0.15 mm, and 0.2 mm are chosen as printing process parameters. The infill pattern and wall thicknesses are not considered due to the design of the prosthetic socket. A broader range of design and printing processes can be explored in future studies by expanding the data range. Fourth, the design and analysis of the prosthetic socket is simplified. The FEA on the prosthetic socket does not consider the connection between the socket and the pylon, which may loosen easily during daily use. More detailed CAD models and analysis are required to predict all possible failures during use. Fifth, prediction results are not compared with physical 3D-printed prosthetic sockets. One future direction can be to perform physical tests on 3D-printed prosthetic sockets and identify the gap. Sixth, only ultimate strength in ISO-10328 P4 loading condition two is selected as the evaluation criterion. In future studies, more criteria can be used to evaluate generated prosthetic designs, such as other strength standards, fatigue resistance, and ventilation. Seventh, the case study is conducted inside the laboratory by researchers. Given the uneven baseline of knowledge among different stakeholders, such as patients, 3D-printing specialists, and healthcare providers, creating user-friendly guidelines that fit each group of users is essential to ensure the effective implementation and adoption of the method. Last, the case study is conducted on a 3D-printed generic residual limb. One future research direction is conducting clinical trials and validating the prosthetic socket generated from the proposed method on human patients.

The authors would like to express gratitude to Dr. Nicholas Okumu and Dr. Ngugi Wamuyu for advising this research on the medical aspects and state-of-the-art prosthetics in the Republic of Kenya. The authors also thank Kevin Marimbet and Aggrey Omondi from the Dedan Kimathi University of Technology for assisting with the research. This work was partly funded by the CMU-Africa African Engineering and Technology Network (Afretec) seed grant. Any opinions, findings, or conclusions in this paper are those of the authors and do not necessarily reflect the sponsors’ views.

There are no conflicts of interest.

The data and information that support the findings of this article are freely available.²