Abstract
This case-control study uses computer vision and artificial intelligence to develop a screening tool for detecting spinal muscular atrophy in infants.
Several underlying conditions can lead to a reduction in muscle tone and generalized hypotonia presenting at birth or in early life, the diagnosis of which is mainly clinical by an experienced physician. In recent years, a growing number of neuromuscular diseases can be treated if detected early, particularly spinal muscular atrophy (SMA).1 Our goal is to develop a screening tool for SMA using computer vision and artificial intelligence (AI).
Methods
We collected 1-minute, 2-dimensional (2-D) videos, using a standard video acquisition system, of 25 pediatric patients. All patients were admitted to our intensive care unit between January 1, 2020, and December 31, 2022. Parental written consent was obtained (Comités de protection des personnes Île-de-France VII approval PP 14-052). This case-control study followed the STROBE reporting guideline.
The 2-D pose estimation data were calculated from the acquired videos using the contrast modification method implemented by Taleb et al2 applied to AlphaPose (Shanghai Jiao Tong University, Machine Vision and Intelligence Group). We first extracted from the 2-D posture the infant’s motor dataset, comprising 12 points corresponding to the articulations, 8 segments corresponding to the limbs, and 4 angles corresponding to limb angles (Figure 1). From those, we calculated different linear, angular, and depth of movements and trained different models to detect SMA. We created 8 tabular datasets corresponding to relevant features to measure. From there, we extracted 108 features and 47 412 corresponding values.
Figure 1. Classification Pipeline.
Numerical skeletons were extracted from motricity videos (step 1). Then, relevant features were measured based on the skeleton (step 2). Finally, a classifier (XGBoost, a supervised machine learning model) labeled data either as control motricity or as spinal muscular atrophy (SMA) motricity (step 3).
We used the Shapley additive explanations (SHap) value proposed by Lundberg et al3 to average the effect of 1 variable on all possible combinations of variables. We used Python, version 3.8.10 (Python Software Foundation), SHap Library, version 0.41.0 (Scott Lundberg), and XGBoost Library, version 1.71 (XGBoost Developers) to classify the features and the dataset.4 Additional details are provided in the eMethods in Supplement 1.
Results
Five patients presented with SMA type 1, with 2 copies of the SMN2 gene confirmed by genetic analysis (mean age, 29.2 weeks; median, 34 weeks; range, 6-56 weeks). The control group comprised the remaining 20 patients who had a normal neurologic examination (mean age, 15.6 weeks; median, 12.5 weeks; range, 1-58 weeks).
The depth of movements, which correspond to the possibility that the child had to move in the frontal plane, a feature of SMA, gave the best discrimination between the 2 groups, with a mean (SD) area under the curve of 0.97 (0.02). The SHap method showed the lower limbs as more discriminant than the upper limbs to classify the motricity as SMA. To better explain the results obtained by SHap, we present a pictorial visualization in Figure 2.
Figure 2. Motor Analysis Based on the Depth of Movement.
Colors indicate the mean of values of the quantity of motricity according to depth, and the dot size represents the outcome on the decision process and arrows the mean importance of each symmetry plan according to the Shapley additive explanations value. SMA indicates spinal muscular atrophy.
Discussion
Using computer vision and AI, we were able to teach a computer to identify abnormal mobility of a group of children presenting with SMA from a control group with an accuracy of 97% and at a younger age than usual.5 Using the SHap model, we showed that the AI decision to classify patients was mainly based on lower limb motricity, consistent with the clinical diagnosis of SMA. A limitation of the study was the small number of infants presenting with neurodevelopmental deficits.
Conclusions
With new therapies for SMA on the horizon, and given the need to counsel families properly, the ability to make a clinically difficult diagnosis early becomes increasingly important for mobilizing limited resources and expertise. Our results suggest that it is possible to use computer vision and AI to identify abnormal motricity for SMA as a first procedural step to diagnosis.
eMethods
Data Sharing Statement
References
- 1.Gowda VL, Fernandez-Garcia MA, Jungbluth H, Wraige E. New treatments in spinal muscular atrophy. Arch Dis Child. 2023;108(7):511-517. doi: 10.1136/archdischild-2021-323605 [DOI] [PubMed] [Google Scholar]
- 2.Taleb A, Diop S, Rambaud P, et al. Improve pose estimation model performance with unlabeled data. In: CSCE 2023 Book of Abstracts. American Council of Science & Education; 2023; 62. Accessed January 29, 2024. https://www.american-cse.org/static/CSCE23-book-abstracts-paperID.pdf
- 3.Lundberg S, Lee S-I. A unified approach to interpreting model predictions. arXiv. Preprint posted online November 24, 2017. doi: 10.48550/arXiv.1705.07874 [DOI]
- 4.Rambaud P, Taleb A, Fauches R, Rimmel A, Tomasik J, Bergounioux J. Binary classification vs. anomaly detection on imbalanced tabular medical datasets. In: CSCE 2023 Book of Abstracts. American Council of Science & Education; 2023; 61. Accessed January 29, 2024. https://www.american-cse.org/static/CSCE23-book-abstracts-paperID.pdf
- 5.Mercuri E, Finkel RS, Muntoni F, et al. ; SMA Care Group . diagnosis and management of spinal muscular atrophy: Part 1: recommendations for diagnosis, rehabilitation, orthopedic and nutritional care. Neuromuscul Disord. 2018;28(2):103-115. doi: 10.1016/j.nmd.2017.11.005 [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
eMethods
Data Sharing Statement
