Feature Engineering for the Prediction of Scoliosis in 5q‐Spinal Muscular Atrophy

Tu‐Lan Vu‐Han, Vikram Sunkara, Rodrigo Bermudez‐Schettino, Jakob Schwechten, Ryan Runge, Carsten Perka, Tobias Winkler, Sebastian Pokutta, Claudia Weiß, Matthias PumbergerJournal of Cachexia Sarcopenia and Muscle 2024 · DOI

TL;DR

This paper tackles the challenge of predicting scoliosis in patients with 5q-Spinal Muscular Atrophy (SMA), a rare and treatable disease. By employing expert domain knowledge for feature engineering and utilizing a Random Forest classifier, the researchers developed a model that achieves 82% accuracy in predicting scoliosis, offering a data-driven approach to improve patient care in this evolving therapeutic landscape.

What Problem Does This Paper Address?

Spinal Muscular Atrophy (SMA) is a rare genetic neuromuscular disease that, despite recent advancements in disease-modifying therapies, presents ongoing challenges for clinicians. One significant complication is scoliosis, which affects a high percentage of SMA patients and is closely linked to disease severity. The introduction of new treatments has created a more dynamic and heterogeneous patient population, making it difficult to rely solely on historical data and traditional classification methods for treatment decisions. This paper aims to address this by developing a predictive model for scoliosis onset in SMA patients, leveraging machine learning and expert knowledge to identify key predictive markers.

Methods

The core of this research lies in its data-driven approach, specifically focusing on feature engineering and machine learning.

  • Feature Engineering: The researchers combined expert domain knowledge with data analysis to create meaningful features from raw patient data. This involved grouping variables related to interventions, devices, orthosis, ventilation, contractures, and motor milestones. They then transformed these variables into numerical or binary formats and aggregated them to create more informative features. A notable example is the "contracture score," which incorporates both the location and severity of contractures, a more nuanced approach than historical methods.
  • Feature Selection: To identify the most relevant features for scoliosis prediction, the team employed correlation analysis and Predictive Power Score (PPS) analysis. PPS is highlighted as a valuable tool for its ability to detect both linear and non-linear relationships, offering insights into potential causality. Expert domain knowledge was crucial in prioritizing features with established clinical relevance.
  • Machine Learning Model: A Random Forest Classifier was chosen for its robustness and interpretability, especially with smaller datasets. The model was trained and evaluated using cross-validation, with performance measured by accuracy and the Area Under the Receiver Operating Characteristic Curve (ROC AUC).

Datasets

The study utilized a newly compiled SMAScoliosis data set comprising data from 84 genetically confirmed 5q-SMA patients. This dataset included 1304 visits and over 360 variables collected through routine clinical assessments. The data was sourced from the SMArtCare study, a large European SMA data registry, ensuring a degree of standardization. The dataset was enriched with paraclinical data and imaging information (PACS).

Key Insights and Ideas

  • Feature Engineering is Crucial for Rare Diseases: The paper strongly advocates for feature engineering, especially in rare disease contexts where data is scarce. They demonstrate that well-engineered features can significantly outperform individual raw variables in predictive power.
  • Dynamic Markers Trump Static Ones: Unlike traditional SMA classification based on static markers like age of onset or SMN2 copy number, this study found that dynamic, easily assessed progression markers (e.g., functional motor scores, ventilation use, contractures) are more predictive of scoliosis. This aligns with the evolving understanding of SMA in the era of new therapies.
  • PPS as a Powerful Tool: The Predictive Power Score (PPS) is presented as a valuable alternative or complement to traditional correlation analysis, offering a more nuanced view of feature relationships and potential causality.
  • Holistic Patient View: The model's success hinges on integrating various aspects of patient care, including growth parameters, functional assessments, and the use of assistive devices and medical interventions.

Key Results

The developed Random Forest model achieved:

  • Accuracy: 0.82
  • Average ROC AUC: 0.87

Key features identified as strong predictors of scoliosis included:

  • Growth Parameters: Age, height, weight, BMI, head circumference.
  • Disease Progression Markers: Hammersmith Functional Motor Scale Expanded (HFMSE) score (PPS=0.27), 6-Minute Walk Test (6MWT) scores (PPS=0.44).
  • Intervention/Support Features: Use of assistive devices (PPS=0.39), ventilation (PPS=0.16), presence of gastric tubes (PPS=0.35), contractures (PPS=0.17).

Limitations and Future Work

  • Data Set Representation: The SMAScoliosis data set overrepresents SMA Type I patients, leading to a skewed age distribution. This might result in "unknown" scoliosis labels for younger patients who are not yet eligible for radiography.
  • Radiograph Limitations: The reliance on spinal radiographs, while the gold standard, is limited by radiation exposure, especially in younger patients. Chest radiographs were used as a secondary method, but may have limitations in sensitivity.
  • Minor Scoliosis Detection: The study acknowledges that minor scoliosis might have been missed due to the sensitivity limitations of clinical exams and chest radiographs.
  • Future Work: The authors suggest future studies could explore a multi-class prediction approach that includes "unknown" scoliosis labels. They also emphasize the need for consistent documentation of relevant progression markers to further refine data-driven models and update SMA classification. The increasing implementation of newborn screening for SMA will also necessitate adaptive approaches to capture evolving disease trajectories.

Summary updated

Abstract

The introduction of disease-modifying SMA therapies, followed by the implementation of SMA in newborn screenings, has presented physicians with never-seen patients. We used feature engineering tools to overcome one of the main challenges when using data-driven approaches in rare disease data sets. Through predictive modelling of this data, we defined disease progression markers, which are easily assessed during patient visits and can help anticipate scoliosis onset. This highlights the importance of progressive features in the drug-induced revolution of this rare disease and further supports the ongoing efforts to update the SMA classification. We advocate for the consistent documentation of relevant progression markers, which will serve as a basis for data-driven models that physicians can use to update their best standard-of-care practices.
BibTeX
@article{vuhan2024feature,
  author = {Tu‐Lan Vu‐Han and Vikram Sunkara and Rodrigo Bermudez‐Schettino and Jakob Schwechten and Ryan Runge and Carsten Perka and Tobias Winkler and Sebastian Pokutta and Claudia Weiß and Matthias Pumberger},
  title = {{Feature Engineering for the Prediction of Scoliosis in 5q‐Spinal Muscular Atrophy}},
  journal = {Journal of Cachexia Sarcopenia and Muscle},
  year = {2024},
  doi = {10.1002/jcsm.13599},
}
Figure