Mattia Perrone
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Mattia Perrone has a total of 23 co-authors. Here are the top co-authors by number of shared publications:
Scott Simmons
Rush University Medical Center, Chicago, USA
4 shared publicationsJohn Martin
Rush University Medical Center, Chicago, USA
3 shared publicationsPhilip Malloy
Rush University Medical Center, Chicago, USA
3 shared publicationsShane J. Nho
Section of Young Adult Hip Surgery, Division of Sports Medicine, Department of Orthopedic Surgery, Rush Medical College of Rush University, Rush University Medical Center, Chicago, IL, USA
3 shared publications- John T. Martin
Department of Orthopedic Surgery, Rush University Medical Center, Chicago, IL, USA
2 shared publications - Philip Malloy
Rush University Medical Center
2 shared publications
Latest Publications
- Open Access
Psychological factors contribute more to chronic low back pain than spine pathology: An LLM-based analysis of radiology reports
Background: Chronic low back pain (cLBP) is the leading global cause of disability, yet anatomical, psychological, and socioeconomic determinants are typically studied in isolation, with limited large-scale integration of imaging and biopsychosocial data. We linked imaging findings with demographic, anthropometric, and biopsychosocial characteristics to determine their relative contributions to cLBP. Methods: We retrospectively analyzed 1,480 Chicago-based patients with lumbar spine MRI between 3/2014 and 3/2024. cLBP was as two or more LBP diagnoses recorded six months apart. Spinal pathologies (stenosis, disc disease, facet arthropathy and spondylolisthesis) were extracted from radiology reports using a large language model (GPT-4) and benchmarked against a regex (rule-based text search) pipeline. Multivariable logistic regression evaluated associations between cLBP diagnosis and demographic, anthropometric, socioeconomic, psychological, and anatomical determinants. A parallel linear regression modeled overall anatomical pathology burden. Results: The LLM pipeline achieved an F1-score of 0.98 versus 0.92 for regex. cLBP diagnosis was most strongly associated with depression (OR=2.69, p<0.001) and anxiety (OR=2.31, p<0.001), with smaller contributions from spinal stenosis (OR=1.07, p=0.006) and spondylolisthesis (OR=1.15, p=0.039). Leave-one-domain-out analysis showed psychological factors were most impactful (ΔAIC=120.1), followed by anatomical (ΔAIC=37.0), demographic/anthropometric (ΔAIC=21.8), and socioeconomic (ΔAIC=3.0). Overall anatomical burden was associated with age, male sex, higher BMI, and Non-Hispanic White race, with no associations with socioeconomic or psychological variables. Conclusions: Factors associated with diagnosed cLBP and those associated with structural spinal pathology are largely dissociated. Psychological comorbidities show the strongest associations with cLBP, while anatomical burden is mostly driven by demographics, reinforcing the psychosocial nature of cLBP.
Open AccessPsychological factors contribute more to chronic low back pain than spine pathology: An LLM-based analysis of radiology reports
Background: Chronic low back pain (cLBP) is the leading global cause of disability, yet anatomical, psychological, and socioeconomic determinants are typically studied in isolation, with limited large-scale integration of imaging and biopsychosocial data. We linked imaging findings with demographic, anthropometric, and biopsychosocial characteristics to determine their relative contributions to cLBP. Methods: We retrospectively analyzed 1,480 Chicago-based patients with lumbar spine MRI between 3/2014 and 3/2024. cLBP was as two or more LBP diagnoses recorded six months apart. Spinal pathologies (stenosis, disc disease, facet arthropathy and spondylolisthesis) were extracted from radiology reports using a large language model (GPT-4) and benchmarked against a regex (rule-based text search) pipeline. Multivariable logistic regression evaluated associations between cLBP diagnosis and demographic, anthropometric, socioeconomic, psychological, and anatomical determinants. A parallel linear regression modeled overall anatomical pathology burden. Results: The LLM pipeline achieved an F1-score of 0.98 versus 0.92 for regex. cLBP diagnosis was most strongly associated with depression (OR=2.69, p<0.001) and anxiety (OR=2.31, p<0.001), with smaller contributions from spinal stenosis (OR=1.07, p=0.006) and spondylolisthesis (OR=1.15, p=0.039). Leave-one-domain-out analysis showed psychological factors were most impactful (ΔAIC=120.1), followed by anatomical (ΔAIC=37.0), demographic/anthropometric (ΔAIC=21.8), and socioeconomic (ΔAIC=3.0). Overall anatomical burden was associated with age, male sex, higher BMI, and Non-Hispanic White race, with no associations with socioeconomic or psychological variables. Conclusions: Factors associated with diagnosed cLBP and those associated with structural spinal pathology are largely dissociated. Psychological comorbidities show the strongest associations with cLBP, while anatomical burden is mostly driven by demographics, reinforcing the psychosocial nature of cLBP.
- 27 September 2025
- Published by Springer Nature in Annals of Biomedical Engineering
- Vol. 54 (1) , 153-162
Gait-to-Contact (G2C): A Novel Deep Learning Framework to Predict Total Knee Replacement Wear from Gait Patterns
Purpose Total knee replacement (TKR) is the most common inpatient surgery in the US. Studies leveraging finite element analysis (FEA) models have shown that variability of gait patterns can lead to significant variability of wear rates in TKR settings. However, FEA models can be resource-intensive and time-consuming to execute, hindering further research in this area. This study introduces a novel deep learning-based surrogate modeling approach aimed at significantly reducing computational costs and processing time compared to traditional FEA models. Methods A published method was used to generate 314 variations of ISO14243-3 (2014) anterior/posterior translation, internal/external rotation, flexion/extension, and axial loading time series, and a validated FEA model was used to calculate linear wear distribution on the polyethylene liner. A deep learning model featuring a transformer-CNN based encoder–decoder architecture was trained to predict linear wear distribution using gait pattern time series as input. Model performance was evaluated by comparing the deep learning and FEA model predictions using metrics such as mean absolute percentage error (MAPE) for relevant geometric features of the wear scar, structural similarity index measure (SSIM), and normalized mutual information (NMI). Results The deep learning model significantly reduced the computational time for generating wear predictions compared to FEA, with the former training and inferring in minutes, and the latter requiring days. Comparisons of deep learning model wear map predictions to FEA results yielded MAPE values below 6% for most of the variables and SSIM and NMI values above 0.88, indicating a high level of agreement. Conclusion The deep learning approach provides a promising alternative to FEA for predicting wear in TKR, with substantial reductions in computational time and comparable accuracy. Future research will aim to apply this methodology to clinical patient data, which could lead to more personalized and timely interventions in TKR settings.
- 27 September 2025
- Published by Springer Nature in Annals of Biomedical Engineering
- Vol. 54 (1) , 153-162
Gait-to-Contact (G2C): A Novel Deep Learning Framework to Predict Total Knee Replacement Wear from Gait Patterns
Purpose Total knee replacement (TKR) is the most common inpatient surgery in the US. Studies leveraging finite element analysis (FEA) models have shown that variability of gait patterns can lead to significant variability of wear rates in TKR settings. However, FEA models can be resource-intensive and time-consuming to execute, hindering further research in this area. This study introduces a novel deep learning-based surrogate modeling approach aimed at significantly reducing computational costs and processing time compared to traditional FEA models. Methods A published method was used to generate 314 variations of ISO14243-3 (2014) anterior/posterior translation, internal/external rotation, flexion/extension, and axial loading time series, and a validated FEA model was used to calculate linear wear distribution on the polyethylene liner. A deep learning model featuring a transformer-CNN based encoder–decoder architecture was trained to predict linear wear distribution using gait pattern time series as input. Model performance was evaluated by comparing the deep learning and FEA model predictions using metrics such as mean absolute percentage error (MAPE) for relevant geometric features of the wear scar, structural similarity index measure (SSIM), and normalized mutual information (NMI). Results The deep learning model significantly reduced the computational time for generating wear predictions compared to FEA, with the former training and inferring in minutes, and the latter requiring days. Comparisons of deep learning model wear map predictions to FEA results yielded MAPE values below 6% for most of the variables and SSIM and NMI values above 0.88, indicating a high level of agreement. Conclusion The deep learning approach provides a promising alternative to FEA for predicting wear in TKR, with substantial reductions in computational time and comparable accuracy. Future research will aim to apply this methodology to clinical patient data, which could lead to more personalized and timely interventions in TKR settings.
- Open Access
A CNN Autoencoder for Learning Latent Disc Geometry from Segmented Lumbar Spine MRI
Purpose Low back pain is the world’s leading cause of disability and pathology of the lumbar intervertebral discs is frequently considered a driver of pain. The geometric characteristics of intervertebral discs offer valuable insights into their mechanical behavior and pathological conditions. In this study, we present a convolutional neural network (CNN) autoencoder to extract latent features from segmented disc MRI. Additionally, we interpret these latent features and demonstrate their utility in identifying disc pathology, providing a complementary perspective to standard geometric measures. Methods We examined 195 sagittal T1-weighted MRI of the lumbar spine from a publicly available multi-institutional dataset. The proposed pipeline includes five main steps: 1) segmenting MRI, 2) training the CNN autoencoder and extracting latent geometric features, 3) measuring standard geometric features, 4) predicting disc narrowing with latent and/or standard geometric features and 5) determining the relationship between latent and standard geometric features. Results Our segmentation model achieved an IoU of 0.82 (95% CI: 0.80–0.84) and DSC of 0.90 (95% CI: 0.89–0.91). The minimum bottleneck size for which the CNN autoencoder converged was 4×1 after 350 epochs (IoU of 0.9984 - 95% CI: 0.9979–0.9989). Combining latent and geometric features improved predictions of disc narrowing compared to using either feature set alone. Latent geometric features encoded for disc shape and angular orientation. Conclusions This study presents a CNN-autoencoder to extract latent features from segmented lumbar disc MRI, enhancing disc narrowing prediction and feature interpretability. Future work will integrate disc voxel intensity to analyze composition.
Open AccessA CNN Autoencoder for Learning Latent Disc Geometry from Segmented Lumbar Spine MRI
Purpose Low back pain is the world’s leading cause of disability and pathology of the lumbar intervertebral discs is frequently considered a driver of pain. The geometric characteristics of intervertebral discs offer valuable insights into their mechanical behavior and pathological conditions. In this study, we present a convolutional neural network (CNN) autoencoder to extract latent features from segmented disc MRI. Additionally, we interpret these latent features and demonstrate their utility in identifying disc pathology, providing a complementary perspective to standard geometric measures. Methods We examined 195 sagittal T1-weighted MRI of the lumbar spine from a publicly available multi-institutional dataset. The proposed pipeline includes five main steps: 1) segmenting MRI, 2) training the CNN autoencoder and extracting latent geometric features, 3) measuring standard geometric features, 4) predicting disc narrowing with latent and/or standard geometric features and 5) determining the relationship between latent and standard geometric features. Results Our segmentation model achieved an IoU of 0.82 (95% CI: 0.80–0.84) and DSC of 0.90 (95% CI: 0.89–0.91). The minimum bottleneck size for which the CNN autoencoder converged was 4×1 after 350 epochs (IoU of 0.9984 - 95% CI: 0.9979–0.9989). Combining latent and geometric features improved predictions of disc narrowing compared to using either feature set alone. Latent geometric features encoded for disc shape and angular orientation. Conclusions This study presents a CNN-autoencoder to extract latent features from segmented lumbar disc MRI, enhancing disc narrowing prediction and feature interpretability. Future work will integrate disc voxel intensity to analyze composition.
- 4 February 2025
- Published by SAGE Publications in Proceedings of the Institution of Mechanical Engineers, Part H: Journal of Engineering in Medicine
- Vol. 239 (2) , 202-211
Synthetic data generation in motion analysis: A generative deep learning framework
Generative deep learning has emerged as a promising data augmentation technique in recent years. This approach becomes particularly valuable in areas such as motion analysis, where it is challenging to collect substantial amounts of data. The objective of the current study is to introduce a data augmentation strategy that relies on a variational autoencoder to generate synthetic data of kinetic and kinematic variables. The kinematic and kinetic variables consist of hip and knee joint angles and moments, respectively, in both sagittal and frontal plane, and ground reaction forces. Statistical parametric mapping (SPM) did not detect significant differences between real and synthetic data for each of the biomechanical variables considered. To further evaluate the effectiveness of this approach, a long-short term model (LSTM) was trained both only on real data (R) and on the combination of real and synthetic data (R&S); the performance of each of these two trained models was then assessed on real test data unseen during training. The principal findings included achieving comparable results in terms of nRMSE when predicting knee joint moments in the frontal (R&S: 9.86% vs R: 10.72%) and sagittal plane (R&S: 9.21% vs R: 9.75%), and hip joint moments in the frontal (R&S: 16.93% vs R: 16.79%) and sagittal plane (R&S: 13.29% vs R: 14.60%). The main novelty of this study lies in introducing an effective data augmentation approach in motion analysis settings.
- 4 February 2025
- Published by SAGE Publications in Proceedings of the Institution of Mechanical Engineers, Part H: Journal of Engineering in Medicine
- Vol. 239 (2) , 202-211
Synthetic data generation in motion analysis: A generative deep learning framework
Generative deep learning has emerged as a promising data augmentation technique in recent years. This approach becomes particularly valuable in areas such as motion analysis, where it is challenging to collect substantial amounts of data. The objective of the current study is to introduce a data augmentation strategy that relies on a variational autoencoder to generate synthetic data of kinetic and kinematic variables. The kinematic and kinetic variables consist of hip and knee joint angles and moments, respectively, in both sagittal and frontal plane, and ground reaction forces. Statistical parametric mapping (SPM) did not detect significant differences between real and synthetic data for each of the biomechanical variables considered. To further evaluate the effectiveness of this approach, a long-short term model (LSTM) was trained both only on real data (R) and on the combination of real and synthetic data (R&S); the performance of each of these two trained models was then assessed on real test data unseen during training. The principal findings included achieving comparable results in terms of nRMSE when predicting knee joint moments in the frontal (R&S: 9.86% vs R: 10.72%) and sagittal plane (R&S: 9.21% vs R: 9.75%), and hip joint moments in the frontal (R&S: 16.93% vs R: 16.79%) and sagittal plane (R&S: 13.29% vs R: 14.60%). The main novelty of this study lies in introducing an effective data augmentation approach in motion analysis settings.
- Open Access
Synthetic Data Generation in Motion Analysis: A Generative Deep Learning Framework
Generative deep learning has emerged as a promising data augmentation technique in recent years. This approach becomes particularly valuable in areas such as motion analysis, where it is challenging to collect substantial amounts of data. The current study introduces a data augmentation strategy that relies on a variational autoencoder to generate synthetic data of kinetic and kinematic variables. The kinematic and kinetic variables consist of hip and knee joint angles and moments, respectively, in both sagittal and frontal plane, and ground reaction forces. Statistical parametric mapping (SPM) did not detect significant differences between real and synthetic data for each of the biomechanical variables considered. To further evaluate the effectiveness of this approach, a long-short term model (LSTM) was trained both only on real data (R) and on the combination of real and synthetic data (R&S); the performance of each of these two trained models was then assessed on real test data unseen during training. The predictive model achieved comparable results in terms of nRMSE when predicting knee joint moments in the frontal (R&S: 9.86% vs R:10.72%) and sagittal plane (R&S: 9.21% vs R: 9.75%), and hip joint moments in the frontal (R&S: 16.93% vs R:16.79%) and sagittal plane (R&S: 13.29% vs R:14.60%). These findings suggest that the proposed methodology is an effective data augmentation approach in motion analysis settings.
Open AccessSynthetic Data Generation in Motion Analysis: A Generative Deep Learning Framework
Generative deep learning has emerged as a promising data augmentation technique in recent years. This approach becomes particularly valuable in areas such as motion analysis, where it is challenging to collect substantial amounts of data. The current study introduces a data augmentation strategy that relies on a variational autoencoder to generate synthetic data of kinetic and kinematic variables. The kinematic and kinetic variables consist of hip and knee joint angles and moments, respectively, in both sagittal and frontal plane, and ground reaction forces. Statistical parametric mapping (SPM) did not detect significant differences between real and synthetic data for each of the biomechanical variables considered. To further evaluate the effectiveness of this approach, a long-short term model (LSTM) was trained both only on real data (R) and on the combination of real and synthetic data (R&S); the performance of each of these two trained models was then assessed on real test data unseen during training. The predictive model achieved comparable results in terms of nRMSE when predicting knee joint moments in the frontal (R&S: 9.86% vs R:10.72%) and sagittal plane (R&S: 9.21% vs R: 9.75%), and hip joint moments in the frontal (R&S: 16.93% vs R:16.79%) and sagittal plane (R&S: 13.29% vs R:14.60%). These findings suggest that the proposed methodology is an effective data augmentation approach in motion analysis settings.
- Open Access
Gait-to-Contact (G2C) - A novel deep learning framework to predict total knee replacement wear from gait patterns
Background: Total knee replacement (TKR) is the most common inpatient surgery in the US. Studies leveraging finite element analysis (FEA) models have shown that variability of gait patterns can lead to significant variability of wear rates in TKR settings. However, FEA models can be resource-intensive and time-consuming to execute, hindering further research in this area. This study introduces a novel deep learning-based surrogate modeling approach aimed at significantly reducing computational costs and processing time compared to traditional FEA models. Methods: A published method was used to generate 314 variations of ISO14243-3(2014) anterior/posterior translation, internal/external rotation, flexion/extension, and axial loading time series, and a validated FEA model was used to calculate linear wear distribution on the polyethylene liner. A deep learning model featuring a transformer-CNN based encoder-decoder architecture was trained to predict linear wear distribution using gait pattern time series as input. Model performance was evaluated by comparing the deep learning and FEA model predictions using metrics such as mean absolute percentage error (MAPE) for relevant geometric features of the wear scar, structural similarity index measure (SSIM) and normalized mutual information (NMI). Results: The deep learning model significantly reduced the computational time for generating wear predictions compared to FEA, with the former training and inferring in minutes, and the latter requiring days. Comparisons of deep learning model wear map predictions to FEA results yielded MAPE values below 6% for most of the variables and SSIM and NMI values above 0.88, indicating a high level of agreement. Conclusion: The deep learning approach provides a promising alternative to FEA for predicting wear in TKR, with substantial reductions in computational time and comparable accuracy. Future research will aim to apply this methodology to clinical patient data, which could lead to more personalized and timely interventions in TKR settings.
Open AccessGait-to-Contact (G2C) - A novel deep learning framework to predict total knee replacement wear from gait patterns
Background: Total knee replacement (TKR) is the most common inpatient surgery in the US. Studies leveraging finite element analysis (FEA) models have shown that variability of gait patterns can lead to significant variability of wear rates in TKR settings. However, FEA models can be resource-intensive and time-consuming to execute, hindering further research in this area. This study introduces a novel deep learning-based surrogate modeling approach aimed at significantly reducing computational costs and processing time compared to traditional FEA models. Methods: A published method was used to generate 314 variations of ISO14243-3(2014) anterior/posterior translation, internal/external rotation, flexion/extension, and axial loading time series, and a validated FEA model was used to calculate linear wear distribution on the polyethylene liner. A deep learning model featuring a transformer-CNN based encoder-decoder architecture was trained to predict linear wear distribution using gait pattern time series as input. Model performance was evaluated by comparing the deep learning and FEA model predictions using metrics such as mean absolute percentage error (MAPE) for relevant geometric features of the wear scar, structural similarity index measure (SSIM) and normalized mutual information (NMI). Results: The deep learning model significantly reduced the computational time for generating wear predictions compared to FEA, with the former training and inferring in minutes, and the latter requiring days. Comparisons of deep learning model wear map predictions to FEA results yielded MAPE values below 6% for most of the variables and SSIM and NMI values above 0.88, indicating a high level of agreement. Conclusion: The deep learning approach provides a promising alternative to FEA for predicting wear in TKR, with substantial reductions in computational time and comparable accuracy. Future research will aim to apply this methodology to clinical patient data, which could lead to more personalized and timely interventions in TKR settings.
- 10 December 2023
- Published by Wiley in Journal of Orthopaedic Research
- Vol. 42 (5) , 1045-1053
Hip joint contact forces are lower in people with femoroacetabular impingement syndrome during squat tasks
It remains unknown if hip joint forces during squat tasks are altered in people with femoroacetabular impingement syndrome (FAIS). The aim of this study is to compare hip joint forces between people with FAIS and healthy controls during double leg squat and single leg squat tasks and within limbs during a single leg squat task in people with FAIS. Kinematic and kinetic data were collected in eight people with FAIS and eight healthy matched controls using 3D motion capture and force plates. AnyBody Modeling System was used to perform musculoskeletal simulations to estimate hip joint angles, forces, and moments for all participants. Estimates were postprocessed with AnyPyTools and converted into normalized time series to be compared using a 1D statistical nonparametric mapping (SnPM) approach. SnPM with an independent samples t‐test model was used to compare people with FAIS to controls, while a paired samples model was used to compare involved to uninvolved limb in people with FAIS. Patients demonstrated lower proximodistal force compared to controls (p < 0.01) and compared to the uninvolved side (p = 0.01) for single leg squat. The smaller joint contact forces in people with FAIS compared to controls could represent a strategy of reduced muscle forces to avoid pain and symptoms during this high demand task. These findings when combined with imaging data could help assess the severity of FAIS on hip related function during higher demand tasks.
- 10 December 2023
- Published by Wiley in Journal of Orthopaedic Research
- Vol. 42 (5) , 1045-1053
Hip joint contact forces are lower in people with femoroacetabular impingement syndrome during squat tasks
It remains unknown if hip joint forces during squat tasks are altered in people with femoroacetabular impingement syndrome (FAIS). The aim of this study is to compare hip joint forces between people with FAIS and healthy controls during double leg squat and single leg squat tasks and within limbs during a single leg squat task in people with FAIS. Kinematic and kinetic data were collected in eight people with FAIS and eight healthy matched controls using 3D motion capture and force plates. AnyBody Modeling System was used to perform musculoskeletal simulations to estimate hip joint angles, forces, and moments for all participants. Estimates were postprocessed with AnyPyTools and converted into normalized time series to be compared using a 1D statistical nonparametric mapping (SnPM) approach. SnPM with an independent samples t‐test model was used to compare people with FAIS to controls, while a paired samples model was used to compare involved to uninvolved limb in people with FAIS. Patients demonstrated lower proximodistal force compared to controls (p < 0.01) and compared to the uninvolved side (p = 0.01) for single leg squat. The smaller joint contact forces in people with FAIS compared to controls could represent a strategy of reduced muscle forces to avoid pain and symptoms during this high demand task. These findings when combined with imaging data could help assess the severity of FAIS on hip related function during higher demand tasks.