💻 computer science

An Ordinal Multi-learner Ensemble with Hidden-Markov-Model Augmentation for Imbalanced Prediction of Post-Therapy Occupational Performance in Children with Cerebral Palsy

This paper proposes an Ordinal Multi-learner Ensemble (OME) augmented with a Hidden Markov Model generator to effectively predict imbalanced, ordinal post-therapy occupational performance scores in children with cerebral palsy, demonstrating superior accuracy and robustness compared to existing baselines and SMOTE.

Original authors: Md Rafiul Hassan, Tonoya Tahsin, Mohammad Anwar Rahman, Sultana Razia, Muhammad Mehedi Hassan

Published 2026-09-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Md Rafiul Hassan, Tonoya Tahsin, Mohammad Anwar Rahman, Sultana Razia, Muhammad Mehedi Hassan

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Every child with cerebral palsy moves through the world in a unique way, their bodies shaped by a condition that affects muscle control and posture. For the therapists who work with them, the daily challenge is not just to understand where a child is now, but to guess where they will be after months of hard work. Therapy is a resource that is often scarce and expensive, so knowing which child might gain the most from a specific treatment helps doctors set realistic goals and use their time wisely. The tools used to measure progress are not simple yes-or-no checks; they are graded scales, like a ladder with rungs from zero to five, where each step represents a distinct level of independence. The problem is that these ladders are often empty at the top and crowded at the bottom. Most children in a group will end up at the same middle level, while only a few make the big leaps to full independence. This makes it incredibly difficult for computer programs to learn from the data, because the rare, successful outcomes are too few to teach the system how to recognize them, and the programs often just guess the common middle answer to look smart.

A team of researchers set out to build a better way to predict these outcomes for children with cerebral palsy. They gathered records from 362 children treated at a rehabilitation center in Bangladesh, looking at 25 different daily activities, from eating and dressing to walking and playing. For each child, they had a score before therapy started and a score after it ended. The goal was to create a computer model that could look at a child's starting point and predict their final score on these 25 activities. The researchers knew that standard computer models would fail here because they treat every score as a separate, unrelated category, ignoring the fact that a score of four is closer to five than it is to one. They also knew that the rare, high scores were so scarce that a model would likely ignore them entirely.

To solve this, the team built a new system called an Ordinal Multi-learner Ensemble. Instead of relying on a single computer program to make the guess, they trained three different types of programs, each looking at the problem in a slightly different way. One program used a method designed specifically for ordered steps, another used a powerful learning tool that adjusts its own rules as it learns, and the third used a special technique to create new, fake examples of the rare, high-scoring children to help the computer understand what success looks like. These three programs then voted on the final answer, but not by a simple majority. Instead, they used a method that finds the middle ground, ensuring that the final prediction respected the ladder-like nature of the scores. The key to their success was the third program, which used a mathematical approach to generate these new examples. It did not just copy and paste existing data; it created new, realistic profiles of children who improved, filling in the gaps where real data was missing.

When the researchers tested this new system, it performed significantly better than any of the individual programs or older methods used to fix unbalanced data. In tests where the data was split and re-split five times to ensure the results were not just luck, the new system consistently predicted the correct outcome more often than the others. It was particularly good at identifying the children who would make meaningful improvements, which is the most important information for a therapist. The researchers also checked how often the computer made mistakes that would matter in the real world. They found that while many errors were small, about half of the mistakes would have shifted a child's classification from "needs some help" to "fully independent," or vice versa. This showed that even a small error in the number could change the entire therapy plan, making the precision of their new system vital.

The study confirmed that combining different types of learning tools, especially one that can intelligently create examples of rare successes, is the best way to handle this kind of difficult medical data. The researchers did not claim that their tool is a perfect crystal ball, and they noted that it was tested on data from only one center, meaning it needs to be checked in other hospitals before it is used widely. However, the results showed that by respecting the order of the scores and filling in the missing pieces of the puzzle, it is possible to give therapists a much clearer picture of what a child's future might look like. This approach offers a way to turn small, messy groups of patient records into reliable guides for decision-making, helping to ensure that every child gets the right kind of support to reach their full potential.

Technical Summary: An Ordinal Multi-learner Ensemble with Hidden-Markov-Model Augmentation for Imbalanced Prediction of Post-Therapy Occupational Performance in Children with Cerebral Palsy

1. Problem Definition

The study addresses the challenge of predicting post-therapy functional outcomes for children with Cerebral Palsy (CP). The specific task involves forecasting ordinal scores (0–5 scale) for 25 distinct occupational performance activities based solely on pre-therapy data. The problem is characterized by two critical, coupled difficulties:

Severe Class Imbalance: The dataset (362 children) exhibits extreme imbalance, with ratios between the most and least frequent scores reaching up to 223:1 (e.g., in toileting tasks). Standard classifiers tend to predict the dominant class, failing to capture rare but clinically significant improvements.
Ordinal Nature of Data: The target scores are ordered, where the distance between predictions carries clinical meaning (e.g., a prediction of 4 vs. 3 is different from 1 vs. 5). Standard nominal classifiers ignore this structure, and standard regression models do not inherently respect the discrete, bounded nature of the scale.

The authors aim to develop a framework that simultaneously handles severe imbalance and respects the ordinal geometry of the data to minimize errors that cross clinically relevant decision boundaries.

2. Methodology

The authors propose an Ordinal Multi-learner Ensemble (OME) that aggregates three complementary base learners using an ordinal median aggregator.

A. The Base Learners

Ordinal Logistic Regression: A proportional-odds model that explicitly models the cumulative probabilities of the ordinal classes.
Gradient-Boosted Ordinal Regression: An XGBoost regressor trained to predict a continuous value, which is then rounded to the nearest integer score.
HMM-Augmented Gradient-Boosted Classifier: A classifier trained on an augmented dataset. The augmentation is generated by a Hidden Markov Model (HMM) designed specifically for this context:
- Structure: The HMM treats ordinal scores as hidden states.
- Emissions: Each state emits features according to a closed-form diagonal Gaussian distribution ( $\mathcal{N}(\mu_s, \Sigma_s)$ ), where parameters are estimated directly from the training data for each score class.
- Transition Matrix: While a pre-to-post transition matrix is estimated to characterize therapy dynamics (showing that therapy typically moves a child up by 0 or 1 level), it is not used to generate synthetic data in the primary experiments. The generator samples directly from the Gaussian emissions of the target post-therapy states.
- Augmentation Strategy: For each minority score $s$ , the generator creates synthetic samples to balance the class distribution (equalizing counts to the majority class). These samples are constrained to the feasible feature space (e.g., age clipped to 0–18 years, scores rounded to 0–5).

B. The Aggregator
The ensemble combines the predictions of the three learners using an ordinal median. The authors prove that among constant predictions, the median minimizes the total absolute ordinal deviation, aligning the aggregation strategy with the clinical cost of errors.

C. Clinical Decision Bands
To evaluate clinical utility, the 0–5 scale is mapped to three therapy-relevant bands:

Unable: Score 0
Some-ability: Scores 1–3
Independent: Scores 4–5
Performance is evaluated not just on exact match accuracy, but on the frequency of errors that cross these band boundaries.

3. Key Contributions

Proposed Framework: Introduction of the OME, a novel ensemble combining ordinal logistic regression, gradient-boosted ordinal regression, and an HMM-based augmentation strategy, aggregated via an ordinal median.
HMM Generator for Imbalanced Ordinal Data: Development of a generative model with closed-form Gaussian emissions per ordinal state. Unlike complex generative models (GANs/VAEs), this approach requires no iterative training, making it suitable for small clinical datasets.
Rigorous Evaluation: The study employs bootstrap confidence intervals, five repeated stratified splits, ablation studies, and a synthetic-data-quality audit. It compares the OME against nominal baselines, SMOTE, and individual ordinal baselines.
Clinical Decision-Band Analysis: Introduction of a metric that quantifies how often prediction errors cross therapy-relevant boundaries, moving beyond simple accuracy to assess clinical impact.

4. Results

The study was evaluated on 25 occupational performance activities across 362 children.

Performance Metrics: On a reference split, the OME achieved a macro-F1 of 0.684, balanced accuracy of 0.674, and accuracy of 0.726.
Comparison: The OME significantly outperformed the nominal baseline, SMOTE, the HMM-alone model, and individual ordinal baselines.
- Removing the HMM-augmented member from the ensemble significantly reduced performance (macro-F1 dropped by 0.039 across splits), confirming the HMM provides essential complementary signal for minority classes.
- Across five repeated splits, the OME remained the most stable and best-performing method (macro-F1 = 0.670 ± 0.009), whereas individual learners (including ordinal logistic regression) showed higher variance and often failed to beat the nominal baseline on average.
Synthetic Data Quality: The HMM-generated samples were found to be novel (not memorized copies) and faithful to real data distributions. They matched real feature moments with high correlation ( $r \ge 0.996$ ) and were less prone to overfitting than SMOTE samples.
Activity-Level Variation: The OME performed best on mobility and communication tasks. It underperformed on highly skewed self-care tasks (e.g., dressing, toileting) where single Gaussian emissions struggled to model complex minority distributions.
Error Analysis: While 88% of errors were off by only one class, approximately 50–52% of all errors crossed a clinical decision band (e.g., predicting "Independent" when the child is actually "Some-ability"). This highlights that single-step errors can still have significant clinical consequences.

5. Significance and Claims

The paper claims to present the first multi-activity, imbalance-aware, ordinal prediction framework specifically designed for post-therapy occupational performance outcomes in children with CP.

Clinical Utility: The study demonstrates that combining ordinal-aware learners with a generative augmentation strategy is superior to standard balancing techniques (like SMOTE) for this specific domain.
Interpretability: The HMM transition matrix serves as a descriptive characterisation of therapy dynamics, offering clinicians a summary of how children typically progress (e.g., the likelihood of moving up one level).
Modest Scope: The authors acknowledge limitations, including the use of a single-center dataset (requiring future multi-center validation) and the simplification of the HMM emission to a single diagonal Gaussian. They do not claim the model dictates therapy plans but rather provides a tool to forecast likely outcomes to help set realistic goals.
Future Directions: The authors suggest that future work should explore mixture emissions to better model multimodal classes and validate the decision-band thresholds with larger clinician panels.

In conclusion, the OME framework offers a robust, statistically significant improvement in predicting ordinal, imbalanced clinical outcomes, with the HMM-augmented member being critical to its success in handling rare improvement cases.