Development and validation of a weight-loss predictor to assist weight loss management

Alexander Biehl, Mikko S. Venäläinen, Laura Suojanen, Sakris K E Kupila, Aila J. Ahola, Kirsi H. Pietiläinen, Laura L. EloScientific 2023 · DOI

TL;DR

This paper introduces a machine learning framework that predicts long-term weight loss outcomes using only self-reported weight data. By leveraging longitudinal weight tracking, the model can identify individuals at risk of not meeting their goals, enabling targeted interventions and more efficient resource allocation in weight management programs.

What Problem Does This Paper Address?

Overweight and obesity are global health crises associated with numerous comorbidities, placing a significant financial burden on healthcare systems. While weight loss interventions are crucial, adherence is often poor, and weight regain is common. Current prediction models for weight loss success often rely on complex energy balance calculations or baseline data, which can be inaccurate or insufficient for personalized, ongoing management. This study aims to bridge this gap by developing a simpler, more accurate predictive tool that utilizes readily available longitudinal weight data to forecast long-term weight loss success.

Methods

The researchers employed a machine learning approach to develop and validate their weight loss predictor. They trained six different models (logistic regression, linear regression, naive Bayes, support vector classification, support vector regression, and multi-layer perceptron) on a Finnish weight coaching cohort. The models were designed to predict weight change outcomes at 9 months, categorized into either three classes (weight loss, insufficient weight loss, weight gain) or five classes (high/moderate/insufficient weight loss, high/low weight gain).

A key methodological innovation was the use of Dynamic Time Warping (DTW) for clustering the weight change trajectories. DTW is adept at comparing time series data by considering the shape and temporal alignment of the data, rather than just the final values. This allowed for a more nuanced understanding of weight change patterns. The models were trained using various input data combinations, including baseline BMI, weight change at different time points, and DTW distances to cluster means. Logistic regression emerged as the best-performing model.

Datasets

  • Training Data: A Finnish weight coaching cohort (Healthy Weight Coaching - HWC) comprising 327 participants (aged 21-78) with at least 9 months of self-reported follow-up weight data.
  • Validation Data: An independent cohort of overweight UK adults (Oxford Food and Activity Behaviours - OxFAB) with 184 participants. This dataset was used to validate the predictive accuracy of the models developed on the HWC cohort.

Key Insights and Ideas

  • Longitudinal Data is Key: The study strongly suggests that tracking weight changes over time, even for short periods, significantly improves the accuracy of predicting long-term weight loss outcomes compared to relying solely on baseline data.
  • Simplicity Enhances Usability: By focusing on self-reported weight data and avoiding complex energy balance calculations, the developed model is more accessible and practical for both individuals and healthcare professionals.
  • DTW for Pattern Recognition: The application of DTW for clustering weight change trajectories proved effective in capturing nuanced patterns of weight loss and gain, leading to more robust classification of outcomes.
  • Early Prediction is Possible: The models demonstrated a remarkable ability to predict outcomes with reasonable accuracy even with very limited data (e.g., half a month), suggesting that early intervention based on these predictions is feasible.

Key Results

  • High Prediction Accuracy: Logistic regression, identified as the best-performing model, achieved impressive prediction accuracies. For the three-class prediction, accuracies reached over 50% with just half a month of data and up to 97% with 8 months of data in the independent validation. For the five-class prediction, accuracies ranged from 39% (0.5 months) to 89% (8 months).
  • Early Non-Responder Identification: The model could predict individuals likely to not meet their weight loss goals with 64% accuracy at 2 months and 73% at 3 months, enabling timely interventions.
  • Web Application Developed: A user-friendly web application has been created, allowing individuals to input their weight data and receive predictions, demonstrating the practical application of the research.

Limitations and Future Work

  • Data Reliability: The accuracy of the predictions is inherently dependent on the reliability of the self-reported weight data. Inconsistent or infrequent data entry, especially in the early stages of an intervention, can reduce prediction accuracy.
  • Behavioral Changes: The model does not account for potential changes in user behavior that might influence weight loss, which could lead to prediction inaccuracies.
  • Future Work: The authors suggest expanding the model to predict weight loss outcomes after 12 months by incorporating more training and test data. They also propose exploring additional data sources as input to further enhance prediction accuracy, especially when limited weight entries are available. Testing other machine learning models is also a potential avenue for future research.

Summary updated

Abstract

This study aims to develop and validate a modeling framework to predict long-term weight change on the basis of self-reported weight data. The aim is to enable focusing resources of health systems on individuals that are at risk of not achieving their goals in weight loss interventions, which would help both health professionals and the individuals in weight loss management. The weight loss prediction models were built on 327 participants, aged 21-78, from a Finnish weight coaching cohort, with at least 9 months of self-reported follow-up weight data during weight loss intervention. With these data, we used six machine learning methods to predict weight loss after 9 months and selected the best performing models for implementation as modeling framework. We trained the models to predict either three classes of weight change (weight loss, insufficient weight loss, weight gain) or five classes (high/moderate/insufficient weight loss, high/low weight gain). Finally, the prediction accuracy was validated with an independent cohort of overweight UK adults (n = 184). Of the six tested modeling approaches, logistic regression performed the best. Most three-class prediction models achieved prediction accuracy of > 50% already with half a month of data and up to 97% with 8 months. The five-class prediction models achieved accuracies from 39% (0.5 months) to 89% (8 months). Our approach provides an accurate prediction method for long-term weight loss, with potential for easier and more efficient management of weight loss interventions in the future. A web application is available: https://elolab.shinyapps.io/WeightChangePredictor/ .The trial is registered at clinicaltrials.gov/ct2/show/NCT04019249 (Clinical Trials Identifier NCT04019249), first posted on 15/07/2019.
BibTeX
@inproceedings{biehl2023development,
  author = {Alexander Biehl and Mikko S. Venäläinen and Laura Suojanen and Sakris K E Kupila and Aila J. Ahola and Kirsi H. Pietiläinen and Laura L. Elo},
  title = {{Development and validation of a weight-loss predictor to assist weight loss management}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41598-023-47930-y},
}
Figure