Cerebral palsy (CP) is the most common movement disorder in young children that negatively impacts function, specifically ambulation (Wu et al., 2004). The vast number of potential interventions to improve ambulation and the diverse decision-making processes used makes treatment planning difficult. To assist in surgical treatment planning and strive toward precision medicine, clinicians refer children for an instrumented gait analysis (IGA). Randomized control trials and quasi-experimental studies demonstrated that not only does the inclusion of IGA data alter surgical decision making but also improves post-interventional outcomes (Ferrari et al., 2015, Fonvig et al., 2020, Gouogh and Shortland, 2008, Lofterod and Terjesen, 2008, MacWilliams et al., 2016, Wren et al., 2011, Wren et al., 2013a, Wren et al., 2013b, Wren et al., 2013c). IGA data are helpful in identifying specific abnormalities, however, only a fraction of the large volume of data can feasibly be used in practice. To address this limitation in clinical care, researchers began experimenting with machine learning (ML) techniques to facilitate the decision-making processes and more effectively recommend evidence-based interventions for children with CP (Arnold et al., 2006, Chia et al., 2020, Galarraga et al., 2016, Galarraga et al., 2017a, Galarraga et al., 2017b, Hersh et al., 1997, Hicks et al., 2011, Reinbolt et al., 2009, Schwartz et al., 2013, Sebsadji et al., 2012, Sullivan et al., 1995).

One of the preliminary steps for applying ML strategies to large data sets is dimensionality reduction. The goals of dimensionality reduction are to decrease the number of features representing a sample, in this case a subject, while keeping as much of the information of the original dataset as possible. For example, common applications of dimensionality reduction for IGA data include the production of summary metrics that represent gait quality. There are several traditional summary metrics such as the Normalcy Index (NI) or Gillette Gait Index (GGI) (Schutte et al., 2000), Gait Deviation Index (GDI) (Schwartz and Rozumalski, 2008), and GDI-kinetics, used describe gait quality. NI is a widely validated metric, particularly for individuals with CP. However, it is computed based on the parameters chosen by clinicians, and it only includes kinematic parameters. McMulkin and MacWilliams (2008) demonstrated that establishing a reliable NI requires a strong lab-specific control set. To address these limitations, the GDI was introduced. The GDI removes subjectivity and is less sensitive to changes of individual gait parameters by considering the entire kinematic time series across the gait cycle. Moreover, GDI compresses the data (Rozumalski and Schwartz, 2008) showed that a reduced representation of gait features provides analytical scopes for classification of crouch gait patterns. Although the analysis of kinematic data is beneficial (Sutherland and Davis, 1993, Rodda et al., 2004), GDI does not include kinetic data that could provide a better representation of the gait patterns (Ounpuu et al., 1991). Most importantly, the linear transformation methods used by GDI and other previous summary metrics forces a limited selection of measures and may not be able to capture the ubiquitous non-linear relationships among gait variables.

With the past decade, deep learning models have become state-of-the-art in complex decision-making problems in computer vision, speech recognition, and natural language processing. Through deep learning approaches, it is possible to combine a large number of weakly-relevant features to build a coherent understanding of how those features relate – from pixels in images to image understanding, and frequencies in audio to robust speech recognition. Deep learning utilizes multilayer neural networks to capture the relationship between low-level measures and assess high-order properties of the information (LeCun et al., 1989, Oja, 1989). Autoencoders are multilayer neural networks designed to simply re-encode presented information to create a low-dimensional representation (Hinton and Salakhutdinov, 2006) and has become a widely used unsupervised feature learning approach. A hierarchy of progressively higher-level features are extracted from the data creating a lower-dimensional set of features that can be used to improve performance in prediction models (Zabalza et al., 2016).

The purpose of the current work is to propose an autoencoder approach to generate a single summary metric (The Shriners Gait Index (SGI)) to quantify the quality of an individual’s gait. The models are trained using high-quality data at the Shriners Children’s MAC in Chicago from 2004 to 2020, which include seven temporo-spatial features, sixty-four lower extremity kinematics features, and forty-three lower extremity kinetics features - a total of 114 features. With the large amount of data available, we can test how well the autoencoder performs in compressing the data compared to a traditional approach. To demonstrate validity of the summary metric (face and concurrent), SGI performance will be compared to an existing summary metric (specifically GDI).