TeenGrowth: Individualized Estimations of Weight-related Risk and Recovery Metrics for Young People with Eating Disorders

Abstract

Objective:

While weight restoration and/or stabilization is crucial for successful treatment and sustained recovery from restrictive EDs, it is often challenging to define an individual’s expected healthy body weight. This paper introduces the TeenGrowth package and its web-based application, designed to calculate and forecast predicted body mass index (BMI) and weight across adolescence.

Method:

TeenGrowth includes functions for data cleaning, predicted BMI z-score and BMI calculations, and growth forecasting. The accompanying Shiny web application provides a user-friendly interface, enabling the identification of predicted weights for individuals. Through a series of 30 computer-simulated datasets for 1100 individuals (1000 ‘healthy’ and 100 ‘ED’), the package’s options for predictive models are evaluated.

Results:

Simulation results highlight potential for use in ED screening and treatment and guide users on modeling options. Prediction of adolescent BMI was more accurate for TeenGrowth models, specifically Mean pre-ED BMIz, Most Recent pre-ED BMIz, or the combination of these metrics (Median BMI error for these methods across all simulations = 0.69) when compared to predictions at the 50^th percentile of population-based norms (Median BMI error = 2.15). Aggregated across simulation approaches, results further support optimal accuracy in identifying ED cases when using Mean, Most Recent, or Mean + Most Recent methods (Mean ED case classification accuracy = .86) as compared to use of a population-based metric – 85% of the 50^th percentile BMI (Mean classification accuracy = .61).

Discussion:

The introduction of TeenGrowth represents a first step towards setting reproducible, personalized predicted body weights for young people.

Introduction

Weight restoration is a cornerstone of successful treatment and sustained recovery from anorexia nervosa (AN) and is often recommended for those with other eating disorder (ED) presentations involving significant weight loss. While not a sufficient marker of full recovery, weight restoration and stabilization is an essential component of restrictive ED treatment³. Individuals who gain more weight during treatment for a restrictive ED fare better over time^4,5. Moreover, early and consistent weight gain serves as a beacon of hope, often signaling a positive treatment trajectory^6–8.

Despite the importance of weight restoration and stabilization to ED recovery, pinpointing an individual’s predicted body weight during recovery is often shrouded in a fog of clinical and scientific uncertainty.^9,10 Patients have diverse premorbid growth trajectories, and not all individuals with restrictive EDs present at underweight status. Historically, the field has leaned on weight-based remission and recovery metrics derived from population-based norms, such as restoring weight above underweight status (e.g. BMI > ~ 19) for adults¹¹ or reaching a normative metric of weight based on age and sex during development¹², as navigational tools. Recently, there has been growing recognition that these universal metrics fail to capture the nuances of illness or recovery across the diverse set of individuals presenting with restrictive EDs^13–17. Significant weight loss, a criteria for atypical AN, is not consistently defined,¹³ and is even more difficult to define in adolescent populations, who may fall from their expected growth trajectory without substantial weight loss.^18,19 While there is a rich literature of approaches to identifying treatment goal weights among young people with EDs, the most common methodologies used in research (e.g. 50^th percentile BMI for age, 50^th percentile weight based on a child’s height (McLauren), weight percentile based on individualized height percentile (Moore)) are not personalized based on premorbid BMI percentile/BMIz.^9,12 Recent definitions of atypical AN severity have attempted to individualize metrics by focusing on amount or percentage of weight loss as opposed to population-based metrics,^13,20 but do not account for ongoing expected development in young people. A promising method of capturing deviation from growth, developmental weight suppression, is defined by difference between one’s highest ever and current BMI z-scores, which demonstrates more consistent associations with psychobiological sequalae of EDs compared to raw weight loss.^19,21 While clinical recommendations for ED treatment in young people include determining treatment goal weights individually, based on prior growth curve data,²² there are no current tools to facilitate reproducible calculations of predicted body weights based on personalized BMI growth curve expectations.⁹

Initial clinical research lends support to a personalized approach to estimation of predicted body weights. A study comparing standardized versus individualized calculations revealed that individualized calculations were typically higher.²³ A recent investigation of young people hospitalized for EDs also identified that premorbid BMIz was associated with both admission BMIz and weight loss prior to admission, further highlighting the importance of weight-related screening and treatment metrics that account for an individual’s growth chart history.²⁴ Further, individuals with the highest levels of weight suppression during an ED often experience more severe symptoms and regain more weight over time,^25,26 which underscores the importance of considering individual weight history when estimating predicted body weights. As the true body diversity and variability in growth history among individuals with restrictive EDs is acknowledged, a crucial understanding also emerges -- weight restoration is not a one-size-fits-all problem, and predicted body weight metrics in ED research and treatment should be personalized and reflect expected developmental shifts with age.

While the potential for personalized predicted body weights to enhance care is significant, the field does not currently have standardized and reproducible methods for approaching this task, representing to a key research and clinical need.^9,15 The development of an evidence-based, individualized approach to defining weight-related ED risk and recovery metrics that capitalizes on growth data charted prior to an ED and offers reproducible, data-driven estimations could serve as a vital navigational tool as providers and patients work towards ED recovery. Here I introduce an R package and accompanying web-based application – TeenGrowth. This novel tool to uses an individual’s growth chart datapoints, demographic data, and ED age of onset as input to identify predicted body weights in a reproducible and statistically-driven manner. I provide an overview of the package and application, which will offer researchers a way to empirically test and optimize predicted body weight calculations and providers a tool to visualize expected growth over time. As a secondary aim, simulated data is used to explore how different settings using the TeenGrowth package perform in a cohort of 1000 ‘healthy control’ children and 100 ‘ED’ cases.

Method

TeenGrowth R Package

The TeenGrowth package is designed to facilitate the calculation, analysis, and forecasting of adolescent growth metrics. The package includes computational functions to facilitate cleaning and processing of raw growth chart data, calculation of BMI and BMI z-scores (BMIz), and generation of growth forecasts across adolescent and young adult development. The functions are documented and organized to provide a toolkit for researchers and clinicians working in adolescent growth assessment. The TeenGrowth package also includes several functions to assist with calculation of growth metrics, including conversion of between BMIz, BMI, and weight values, and the ability to add BMIz values to datasets with BMI and age values, and vis versa, using CDC ²⁷ and WHO ²⁸ growth reference charts. Table 1 provides key terms and definitions used in the paper.

Table 1.

Terms and Definitions

Term	Definition
BMIz	BMI z-score, based on population-based BMI percentile reference data across development, calculated within age and sex, available from ages 2–21 years. BMI follows a non-linear trajectory during childhood, necessitating reference data to define BMI percentiles/BMIz across age and sex. Follows a normal distribution centered at 0.
BMI Percentile	Derived from BMI z-score using the cumulative distribution function of a standard normal distribution
BMIz point prediction	A predicted BMIz score which, in the TeenGrowth package, is personalized based on an individual’s prior BMIz data. Point prediction options in the TeenGrowth package include: Mean: Mean BMIz through ED onset Most Recent: Most recent BMIz obtained prior to ED onset Mean + Most Recent: An average of the mean + most recent obtained BMIz values Maximum: The maximum BMIz that an individual reached in their growth chart.
BMI percentile point prediction	A predicted BMI percentile corresponding to BMIz point prediction, derived from the cumulative distribution function of the standard normal distribution.
Predicted BMI	predicted BMI, which is calculated using the following information:  1. A BMIz/BMI percentile point prediction (previously derived in the TeenGrowth Package)  2. Child Age  3. Child Sex While BMIz and BMI percentile point prediction remain stable across age, predicted BMI will rise in accordance with growth curve expectations
Predicted Weight	Predicted body weight – Calculated using the following information:  1. Predicted BMI at a specified age  2. Adult Height (not provided or calculated prior to reaching adult height)1
BMI/Weight/BMIz Prediction Window	A range that represents the upper and lower bounds of predicted BMI, weight, or BMIz. This window may be tailored to reflect a preferred estimation range. Options include:  1. 80% prediction interval2  2. 95% prediction interval2  3. 99% prediction interval2  4. A pre-specified, symmetrical BMIz range (e.g. +/− 0.5 SD).3

Note:

¹Weight restoration plots for short-term treatment planning will provide a predicted body weight based on current height, while long-term predicted weights are not provided until adult height is obtained Table 1 provides terms and definitions related to use of the TeenGrowth package. BMI = Body Mass Index.

²Based on individual variability in premorbid BMIz.

³As an example, with a BMIz point prediction of 0 and range of +/− 0.5, predicted BMI percentile would be the 50^th, with an expected range between the 31^st and 69^th percentile. For a BMIz point prediction of 1 (+/− 0.5), predicted BMI percentile would be the 84^th (range: 69^th-93^rd)

Forecasting Growth Metrics.

The core motivation for the TeenGrowth package is forecasting predicted BMI. To this end, there are functions for forecasting BMIz across age using several methods of point prediction (mean BMIz, most recent BMIz prior to ED onset, maximum BMIz [i.e. aligning with developmental weight suppression¹⁹]). In addition, the package offers options for specifying the width of BMIz prediction intervals, which can either be personalized based on prior variance of a particular individual’s BMIz values or can be set to a standardized window (e.g. +/− 0.5 BMIz units). The package also offers tools for visualizing projected growth data -- including past and projected BMI values, projected weights after adult height is achieved, and weight restoration plans (See Results for example output). An example analysis workflow using example the package might involve the following steps:

1. Data Cleaning: Load raw data and prepare it for forecasting using clean_data.

2. Forecasting: Generate future growth forecasts using forecast_bmi.

3. Visualization: Plot the results with plot_eBMI and plot_weight to visualize growth trends.

Documentation and Usage.

The package and all contents are available at https://github.com/embark-lab/TeenGrowth. A README file provides basic information on use, a package vignette available at https://embark-lab.github.io/TeenGrowth/articles/Vignette.html guides users through the process of cleaning and forecasting growth data.

TeenGrowth Web Application

Overview.

The TeenGrowth web application (https://embark-lab.shinyapps.io/teengrowthapp) was developed to harness the functions developed in the R package through a more user-friendly interface – without users needing knowledge of how to use R or associated packages. This Shiny web application provides users with a tool to aid in the prediction of individualized growth predictions and plan weight restoration for adolescents with EDs.

User Interface.

The main components of the Shiny app UI are as follows:

• Data Input: Users can upload their own data files (.csv or .xlsx) or use provided demonstration data. The uploaded data must contain information related to age, sex, and anthropometric indicators (one of the following: BMIz, BMI, BMI percentile, or a combination of weight and height). The inclusion of additional information will allow for further utility from the application. For instance, the specification of adult height will allow for estimation of predicted body weights after the age at which adult height is reached.

• Data Specification: Users specify the names of columns the uploaded data, such as age, sex, height, and weight. Users also specify the units of measurement for age, height, and weight. As such, multiple data structures are allowed for input (e.g. ages in years, months, or days are allowed; ‘height’, ‘ht’, or ‘ht_in’ would all be acceptable column names for a height column).

• Model Selection: Users can run models on the specified data to generate predictions and visualize predicted BMI and weight over time

• Weight Restoration Planning: Users can plan for weight restoration based on the predicted weights. This tab includes inputs for treatment start date, current height, and other relevant parameters. It generates plots to visualize weight restoration plans.

Simulating statistical approaches to weight-related prediction

The primary aim of the current paper is to introduce functionality of the TeenGrowth Package, with the understanding that further research is needed to (1) specify optimal approaches for BMI/weight prediction with clinical and population-based samples, (2) evaluate and refine usability of the Shiny Application, and (3) develop recommendations for how and when to best apply this tool in treatment settings. As a first step in evaluation of differing approaches to weight-related prediction, I conducted a series of analysis on computer data simulated under varying conditions with a range of forecasting options available from the TeenGrowth package. Full code for replication of the simulation study is available at https://github.com/embark-lab/TeenGrowth_Simulations/tree/main.

Data Generation and Simulation

Longitudinal BMIz score data was fully simulated for a cohort of children to investigate the impact of EDs on growth patterns. The simulation created data for 1000 control children and 100 children with an ED onsetting after age 14 (control: 50% female; ED: 80% female). Research consistently finds higher prevalence of restrictive-spectrum EDs in females, compared to males^29–33. While the ratio of females to males affected by restrictive EDs varies among samples (between 3:1 to 10:1), a 4:1 female-to-male ratio of individuals impacted by ED-related weight loss was chosen for simulations, to be reflective of an expected sex differential in diagnosis within a larger population that is more evenly distributed across sex. A series of 30 simulations were completed with varying parameters, as outlined below. Each child had a series of BMIz score measurements were simulated over an age range from 24 months (2 years) to 240 months (20 years). Measurements were divided such that two-thirds of the data points were taken at random intervals before age 14.0 years, and one-third were taken at random intervals after age 14.0. The number of total data points per child varied, with possible values of 3, 6, 9, 12, and 15. The autocorrelation parameters (rho) used in the simulation were set to 0.7 and 0.9 per year, reflecting moderate-to-high year-over-year autocorrelation (conservative estimates relative to BMI autocorrelation in adult samples).³⁴ The within-child standard deviations for BMIz scores were set to 0.2 and 0.4 across the varying simulations, based on estimations from prior research.³⁵ After simulated training data was produced, data collected prior to age 14 was separately used to predict BMIz from ages 14–20 using the TeenGrowth package. Further Details on the simulation procedure and forecasting are available in Supplemental Methods.

Statistical Analysis:

Forecasts from the training data were compared against the testing data to evaluate the impact of different central values and confidence intervals on BMIz score prediction accuracy. Simulation parameters (number of data points, within-child variability in BMIz, degree of autocorrelation) were compared across forecasting approaches (points of prediction and prediction windows). The effectiveness of these forecasts was assessed for predicting BMIz scores in children with and without EDs. Three primary outcomes were assessed:

1. Median BMI Prediction Error: Evaluated the accuracy of predicted BMI datapoints for healthy control children after age 14, under varying data generation methods and forecasting approaches.

2. Screening Method Utility: Examined potential utility by coding each datapoint in the test set as falling above or below the prediction window and identifying whether the individual was a control or ED participant. This captured true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN). The sensitivity ( $\frac{\mathrm{T}\mathrm{P}}{TP+FN}$) specificity ( $\frac{\mathrm{T}\mathrm{N}}{TN+FP}$), positive predictive value ( $\frac{\mathrm{T}\mathrm{P}}{TP+FP}$), negative predictive value ( $\frac{\mathrm{T}\mathrm{N}}{TN+FN}$), and balanced accuracy ( $\frac{\mathrm{s}\mathrm{e}\mathrm{n}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{i}\mathrm{t}\mathrm{y}+\mathrm{s}\mathrm{p}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{f}\mathrm{i}\mathrm{c}\mathrm{i}\mathrm{t}\mathrm{y}}{2}$) were coded for each simulation and forecasting approach to determine approaches that may aid in identifying which forecasting methods may be optimal as screening metrics – for instance, identifying individuals that might require further ED screening – based on differences between predicted BMI/BMIz and actual BMI/BMIz in adolescence.

Results

Below I provide (1) example output from the TeenGrowth package and (2) simulation study results.

TeenGrowth Package and Application – Example Output

Example output of the TeenGrowth web application is presented below (Figure 1), using predicted BMI point estimates based on the mean of predicted BMIz scores prior to ED onset and a prediction window based on a 95% confidence interval from this mean. The teen in this example had a premorbid BMI above the median weight range and onset of atypical AN around age 12. They lost a significant amount of weight during their ED but never reached an underweight threshold. At age 15.5, she began family-based therapy, with a goal of 0.5–1 pounds per week of weight restoration (estimated time to weight restoration ~10 months). After four months in treatment, there is a clear lack of progress with weight restoration (Figure 1c). As early weight gain is a predictor of success in this treatment,^4,6 the patient’s current trajectory could signal to providers that an adjustment in treatment strategy (e.g. adding nutritional support; enhancing mealtime supervision; transition to a different treatment approach) could be considered. An additional example based on an individual with AN whose premorbid weight trended at a lower BMIz is presented in Supplemental Figure 1. Predicted weight tables for both teens, provided by the TeenGrowth Package, are available in Supplemental Table 1. Similar output can be provided by the R Package, for researchers, and via the Shiny application, for non-technical users, with the ability for users to specify various methods for BMI point prediction and prediction interval size.

Figure 1. Example output from an individual with atypical AN. Teen in this example had an ED onset ~age 15.

Note. This teen’s premorbid mean BMIz was at the 80^th percentile, and they lost a significant amount of weight during their ED without reaching underweight status. They began treatment immediately in an outpatient setting, with a goal of 0.5–1 pounds per week towards weight restoration, though there is a lack of early treatment progress. Panels a-b present predicted BMI and weight based on premorbid mean BMIz with a 95% prediction interval. Panels C-D show current treatment progress including projected weight restoration (Panel C) and BMI percentile (Panel D). Panel E provides summary results. Similar output is available through both the TeenGrowth package and Shiny application.

Simulation Study

Accuracy of Expected BMI:

Training data (prior to age 14) was used to forecast points after age 14, and results were compared with testing data matched for timepoint within each child. The absolute difference between forecast and testing data evaluated predicted BMI forecast accuracy. Full results are presented in Figure 2 and available in Supplemental Table 2. Both the within-child standard deviation of BMIz and the BMIz autocorrelation impacted estimation error. Predictions were quite accurate with low within-child variability (SD = 0.2 BMIz units), with median BMI prediction less than one BMI point (i.e. < 6lb for an average height individual) for all individualized estimation methods. Mean, Most Recent, and Mean + Most Recent methods preformed similarly, and an increase in the number of datapoints improved estimations for these methods when autocorrelation was moderate. Median BMI error for the Mean, Most Recent, and Mean + Most Recent methods when within-child SD was 0.2 was 0.47, which is equivalent to an error of only ~3lb/1.4kg for an individual 65 inches tall. When within-child SD was higher (0.4), the median BMI error for these methods was 0.98, equivalent to ~ 6lb in the same individual. On the other hand, when using the 50^th percentile BMI estimation for all individuals (Population Median), the average BMI error was 2.04 when within-child SD was 0.2 and 2.21 when within-child SD was 0.4, equivalent to a ~12–14lb error for an adolescent at median height. The Maximum method outperformed the Population Median method but was not as robust as other individualized methods.

Figure 2. Median Difference between Expected and Observed BMI in the Test Data (> age 14) Based on Number of Training Datapoints, Within-Child Standard Deviation Prediction Points, and Autocorrelation, within the Control Sample.

Note. ‘Population Median’ refers to the 50^th percentile BMI based on age- and sex-referenced CDC growth charts. One BMI point is equal to ~ 6lb/1.4 kg for an individual of average height.

Prediction of ED cases.

To examine potential utility of the TeenGrowth package for screening purposes in this simulation study, we evaluated sensitivity, specificity, positive predictive value, negative predictive value (see Supplemental Figure 2 for all metrics), and balanced accuracy (See Figure 3) all measured on a 0–1 scale; with higher numbers indicating better accuracy of different approaches in capturing ED cases. Overall, the Mean, Mean + Most Recent, and Most Recent models were all quite accurate. The Maximum model was less accurate in most cases, and the Population Median method was substantially less accurate than other methods. The Population Median method demonstrated low sensitivity – indicating that it was not effective at capturing the majority of ED cases. When within-child BMIz SD was moderate (0.4), accuracy of individualized methods was diminished, though still more accurate relative to the Population Median Method. Average accuracy within simulation methods (collapsing across N Datapoints and Autocorrelation metrics) are presented in Table 2, and full results for all accuracy metrics are presented in Supplemental Table 3.

Figure 3. Balanced accuracy of Different Forecasting Approaches in Predicting ED vs. Control Status in Individuals after age 14.

Note. ‘Population Median’ refers to the 50^th percentile BMI based on age- and sex-referenced CDC growth charts

Table 2.

Mean balanced accuracy by forecast point and within-child standard deviation

Forecast point	Mean Accuracy
Within child SD = 0.2
Max	0.86
Mean	0.89
Mean + Most Recent	0.90
Most Recent	0.90
Population Median	0.61
Within child SD = 0.4
Max	0.79
Mean	0.81
Mean + Most Recent	0.81
Most Recent	0.81
Population Median	0.62

Discussion

Package and Application Strengths

The TeenGrowth package and app present a novel, versatile solution for tracking and predicting BMI and weight changes in youth. Notable features include utility that extends into young adulthood and automatic calculation of BMIz scores from growth chart data and the projection of BMI and weight changes as youth age. Beyond providing point estimates for weight-related recovery metrics, the package also offers tailored prediction ranges, which, recognizing that point estimates will not be perfectly accurate, offer a range that might represent a lower bound of where BMI percentile, BMI, and weight might fall. The package also holds promise for improving ED screening in more general pediatric settings by identifying adolescents whose growth falls outside their individualized forecasted BMIz/BMI percentile range. Regarding how this tool could eventually impact ED research and clinical practice, a survey of child and adolescent ED treatment providers noted that, while use of historical growth curves is common, less than half of clinicians reported their use in developing estimations of treatment goal weights and lack of consensus in how to estimate goal weights in the field.⁹ This study sampled from ED organizations – lack of integration of growth chart metrics in defining predicted body weights may be even more pronounced among individuals who do not specialize in ED treatment, but who nonetheless may be front-line providers for ED screening. Clinical recommendations suggest incorporation of historical growth curve data in estimations of body weight in ED treatment;²² however, there is currently no evidence-based method to calculate an expected or predicted body weight that incorporates this data.⁹ While further testing is necessary, TeenGrowth provides (1) a tool for researchers to easily estimate, compare, and ultimately optimize these predictions and models along with (2) a reproducible platform for providers to conduct these calculations for individual patients, which can be further refined with future usability testing.

Current Recommendations for Use

In addition to introducing the package, simulation evaluation allowed for initial evaluation of forecasting approaches. Thirty scenarios were tested, incorporating variations in within-child variance, year-over-year autocorrelation, and the number of data points available for each child. Overall, results from the current simulation study bore some initial fruit regarding usage recommendations:

1. The Mean, Most Recent, or Mean + Most Recent forecasting options are currently recommended as the first line forecasting approaches. All outperformed the Population Median approach for predicting BMI.

2. If year-over-year autocorrelation in BMIz is known to be high (defined here as 0.9), using the most recent BMI datapoint prior to ED onset tends to provide the most accurate BMI estimations.

3. Use of CDC (rather than WHO) reference charts is generally recommended. Users should be mindful that forecasts are reliant on existing growth reference charts, which may not be completely reflective of expected growth trajectories for all children. For instance, reference growth charts are sex-specific, thus, a binary sex variable must be specified for all inputs.

Package and Application Limitations

Further research on TeenGrowth is needed to (1) specify optimal approaches for predicted body weight calculation with both clinical and population-based samples, (2) evaluate and refine usability of the Shiny Application, and (3) develop recommendations for how and when to best apply this tool in treatment settings. More specifically, additional research could clarify a host of questions such as: Is there a single method of prediction that outperforms other methods? Does variability in growth history impact accuracy of individualized predicted body weights? What additional, external, variables could improve prediction accuracy, and how should they be weighted? What are the largest barriers to implementation of this tool? Research on efficacy, feasibility, and acceptability is necessary to determine the extent to which TeenGrowth could benefit key stakeholders.

At present, the accuracy of TeenGrowth is contingent on accurate historical growth data prior to ED onset. Although retrospective recalls can theoretically fill in growth chart gaps, this approach may introduce recall bias and impact reliability. Further, the current iteration of the application requires the age of onset for the ED to be pre-identified for accurate predictions. Additional investigation will clarify best practices for implementation –including how to identify datapoints that should be added to or excluded from an individual’s pre-ED forecasting data and how to best specify ED age of onset.

Simulation Limitations

The simulations, while robust, do not consider all possible parameters for data generation. One parameter that was not included was a BMIz set point that places pressure to return to a highest ever BMIz, which could upwardly bias Maximum BMIz error estimate and downwardly bias the errors of other models. The current approach was chosen as BMIz is expected to remain stable based on the methodologies used in current growth reference charts. Additional testing in real-world data may clarify typical parameter values and further refine recommendations for which forecasting approaches may be optimal defaults. Further, while the simulations aimed to refine the package’s predictive models, external factors like medication, health changes, or pubertal timing on an individual’s growth trajectory, may prove additional challenges for real-world applications. Related to this point, while population-based estimations using median BMIz were incorporated as an initial comparison for assessing prediction accuracy, many eating disorder specialists implement individualized treatment goal weights in clinical practice. Future work could consider benchmarking TeenGrowth estimations of predicted body weights with those of expert clinicians, who may implicitly or explicitly consider additional variables beyond growth chart data when defining treatment goal weights. In time, and with appropriate data to support accurate estimation of the impacts of such variables, the influence of external variables could be incorporated into predictive modeling. Overall, while the simulations provide valuable insights and help in refining the forecasting approaches, ongoing research and validation is vital to evaluate and refine the utility of the package.

In Defense of Improving and Personalizing Predicted Body Weight Estimations

A critic of this highly quantitative endeavor might argue that the ED field should not be developing tools aimed at increasing the accuracy of predicted body weights. They may contend that calculating treatment goal weights is a fool’s errand that will entrench rigidity and focus on weight, undermining a dynamic approach to holistic ED recovery that includes nutritional flexibility, the absence of ED behaviors, and reduced importance of weight and shape.¹⁶ Improving estimation procedures for treatment goal weights need not mean that these estimations should be an exclusive focus of recovery, nor that these estimations are perfectly accurate. The package architecture and data generating mechanisms of simulations produced in this paper reflect that estimation error is natural, that predicted body weights will change across growth and development, and that a moving window with age provides a more robust sense of where weight might settle as compared to a single data point.

The TeenGrowth tool can be viewed a compass - guiding direction without prescribing a fixed route. The utility of this compass will ultimately rely on wise implementation with skilled providers at the helm, complementing rather than dominating medical, behavioral, and psychological metrics. Despite these caveats, the package holds potential for improving ED care. At present, individuals with atypical AN receive delayed, suboptimal care,¹⁴ and clinicians report difficulty estimating treatment goal weights for this population.¹⁵ Further, even among those who reach underweight status and receive an AN diagnosis, premorbid growth trajectories vary substantially. The introduction of a tool which provides individualized predicted body weights based on historical growth metrics has potential to bolster the case for better, more adequate care for those with a restrictive ED onset in higher weight bodies and to improve scalability of individually-defined ED screening and treatment.

Conclusions

In sum, the TeenGrowth package and accompanying web application offer a statistically-driven, individualized approach to estimating predicted body weights. The package’s ability to handle diverse data inputs, provide individualized BMI and weight forecasts, and offer prediction ranges enhances its utility. Future research and real-world validation will further refine its applications, enhance its usability, and broaden its impact. The TeenGrowth package holds promise for improving ED screening, facilitating more accurate estimations of predicted body weight among individuals with diverse growth histories, and fostering greater reproducibility and transparency in the estimation of weight related risk and recovery metrics in ED care.

Public Significance Statement:

The TeenGrowth package offers a promising, individualized approach to estimating weight-related risk and recovery metrics in adolescents with eating disorders. By providing personalized, data-driven predictions, it offers potential to enhance clinical decision-making and treatment outcomes. While further research and validation are needed, this tool holds promise to advance precision medicine in eating disorder recovery.