Machine Learning Prediction of Pulmonary Function in Children With Cystic Fibrosis

This study is a retrospective, cross-sectional, non-interventional secondary analysis of de-identified archived data obtained from children with cystic fibrosis who were evaluated at the Cardiopulmonary Rehabilitation Unit, Department of Physiotherapy and Rehabilitation, Faculty of Health Sciences, Gazi University, between April 2021 and September 2022.

The source data were collected prospectively as part of a previous observational study approved by the Gazi University Ethics Commission (Decision No. 839, December 7, 2020) and prospectively registered at ClinicalTrials.gov (NCT04803643). Written informed consent was obtained from the participants and their parents or legal guardians before the assessments conducted in the source study. No new participant contact, recruitment, assessment, biological sample collection, or intervention will be performed for the present secondary analysis.

Although the parent study included an age- and sex-matched healthy control group, the present prediction analysis will be restricted to the archived records of 31 children with cystic fibrosis who satisfy the prespecified eligibility criteria and have the variables required for the planned analyses.

Pulmonary Function Outcomes

Pulmonary function was assessed using a COSMED spirometer (Class II/Internally Powered Equipment, Italy) in accordance with American Thoracic Society/European Respiratory Society standardization criteria. Measurements were performed in the seated position using a nose clip. Participants completed at least three technically acceptable forced expiratory maneuvers, and the highest acceptable values were recorded.

Forced expiratory volume in one second (FEV1), forced vital capacity (FVC), and peak expiratory flow (PEF), expressed as both absolute and percent-predicted values, will constitute the prespecified pulmonary function prediction targets. A separate prediction model will be developed for each pulmonary function target.

Peripheral Muscle Oxygenation

Peripheral muscle oxygenation was assessed noninvasively using the Moxy Monitor, a portable near-infrared spectroscopy device (Fortiori Design LLC, Minnesota, USA). During the six-minute walk test, the sensor was positioned over the quadriceps femoris muscle of the dominant lower extremity and secured in the same position throughout the test.

Muscle oxygen saturation (SmO2) and total hemoglobin (THb) were monitored continuously during the resting, exercise, and recovery periods. The archived measurements included resting, minimum, maximum, and recovery SmO2 and THb values. Changes between resting, exercise, and recovery values will also be calculated when the required measurements are available. These muscle oxygenation variables will be treated as primary candidate predictors.

Functional Exercise Capacity and Oxygen Saturation Response

Functional exercise capacity was assessed using the six-minute walk test in accordance with American Thoracic Society procedures. Participants were instructed to walk, without running, as far as possible along a predetermined level corridor for six minutes. The total distance walked was recorded in meters.

Heart rate, blood pressure, respiratory rate, peripheral oxygen saturation, perceived dyspnea, and leg fatigue were assessed before and after the test. Peripheral oxygen saturation was monitored throughout the test, and resting, minimum, end-test, and recovery SpO2 values were recorded. Exercise-related changes in SpO2 will be derived from these archived measurements when appropriate.

Perceived dyspnea and fatigue were evaluated using the Modified Borg Scale, ranging from 0 to 10, with higher scores indicating greater symptom severity. Six-minute walk distance, SpO2 responses, and symptom scores will be evaluated as additional candidate predictors.

Respiratory Muscle Strength

Respiratory muscle strength was assessed using a portable electronic mouth-pressure measurement device in accordance with ATS/ERS recommendations. Maximal inspiratory pressure was measured from a lung volume close to residual volume, whereas maximal expiratory pressure was measured from a lung volume close to total lung capacity.

Participants performed at least three technically acceptable and reproducible maneuvers with verbal encouragement. The highest acceptable maximal inspiratory pressure and maximal expiratory pressure values were recorded in cmH2O. MIP and MEP values were also expressed as percentages of age- and sex-predicted values when available. Absolute and percent-predicted MIP and MEP measurements will be evaluated as additional candidate predictors.

Inspiratory Muscle Endurance

Inspiratory muscle endurance was assessed using an incremental threshold-loading test with the POWERbreathe Wellness device (POWERbreathe International Ltd., Birmingham, United Kingdom). The test was initiated at 20% of the participant's maximal inspiratory pressure. The threshold load was subsequently increased to 40%, 60%, 80%, and 100% of MIP at two-minute intervals.

Participants were instructed to continue breathing against the imposed resistance at each load. The test was terminated if the participant failed to complete three consecutive breaths, could not maintain the required breathing pattern, could not tolerate the load, requested termination, or met a clinical stopping criterion.

The archived respiratory muscle endurance variables included total test duration, number of breaths completed, highest successfully sustained load, and the highest load expressed as a percentage of MIP. These variables will be considered additional candidate predictors.

Static Postural Balance

Static postural balance was assessed using the Biodex Balance System. Testing was performed under four sensory conditions: eyes open on a firm surface, eyes closed on a firm surface, eyes open on a soft surface, and eyes closed on a soft surface.

Each condition was assessed three times for 20 seconds, with a 10-second rest interval between repetitions. Overall Stability Index, Anteroposterior Stability Index, and Mediolateral Stability Index values were recorded. Higher stability index values indicate greater postural sway and poorer postural stability. The stability indices obtained under the different sensory conditions will be treated as primary candidate predictors.

Objectively Measured Physical Activity

Physical activity was assessed objectively using the SenseWear Armband multisensor activity monitor (BodyMedia Inc., Pittsburgh, Pennsylvania, USA). The monitor was positioned over the triceps region of the nondominant upper arm and was worn continuously for three days, except during bathing.

The archived physical activity measurements included daily step count, total energy expenditure, active energy expenditure, physical activity duration, average metabolic equivalent, time spent at different physical activity intensities, sedentary or lying time, and sleep duration. Measurements from valid monitoring days were averaged and processed using SenseWear Software version 7.0. The physical activity variables available in the archived dataset will be evaluated as additional candidate predictors.

Cystic Fibrosis-Specific Quality of Life

Cystic fibrosis-specific health-related quality of life was assessed using the validated Turkish version of the Cystic Fibrosis Questionnaire-Revised. The age-appropriate child, adolescent, or parent-report version was administered in the source study.

The CFQ-R evaluates domains including physical functioning, emotional functioning, social or school functioning, body image, eating problems, treatment burden, health perceptions, weight perceptions, respiratory symptoms, and digestive symptoms, depending on the age-specific questionnaire version. Each domain is transformed to a score ranging from 0 to 100, with higher scores indicating better health-related quality of life. Individual domain scores, rather than an unvalidated overall total score, will be evaluated as additional candidate predictors.

Anthropometric, Nutritional, and Clinical Variables

Archived demographic and anthropometric variables included age, sex, body weight, height, body mass index, weight-for-age Z score, height-for-age Z score, and BMI-for-age Z score. BMI was calculated as body weight in kilograms divided by height in meters squared.

Disease-related clinical characteristics included age at diagnosis, disease duration, pulmonary exacerbations during the preceding year, hospital visits and admissions, medication use, treatment burden, smoking or environmental exposure, and other relevant clinical characteristics available in the archived dataset. These anthropometric, nutritional, demographic, and clinical variables will be evaluated as additional candidate predictors.

Statistical and Machine-Learning Analysis

Separate machine-learning regression models will be developed for the prespecified pulmonary function targets. Peripheral muscle oxygenation and postural balance variables will be treated as the primary candidate predictors. Functional exercise capacity, respiratory muscle strength and endurance, exercise-related oxygen saturation, physical activity, quality of life, anthropometric and nutritional indicators, and disease-related clinical characteristics will be treated as additional candidate predictors.

Categorical variables will be encoded in a form suitable for the selected models. Continuous variables will be examined for implausible values, outliers, distributional characteristics, and strong correlations. The extent and pattern of missing data will be described. Depending on the amount and pattern of missingness, complete-case analysis or an appropriate imputation procedure will be used.

Because of the limited sample size and the relatively large number of candidate predictors, model complexity and the number of variables entering an individual model will be restricted. Parsimonious and regularized regression approaches suitable for small datasets will be prioritized. Alternative low-complexity nonlinear models may be examined only when supported by the available data.

A fixed training-test split will not be used because it would produce very small and potentially unstable subsets. Internal validation will instead be based on cross-validation. All data-dependent preprocessing procedures, including imputation, standardization, variable selection, and hyperparameter tuning, will be performed using the training data within the cross-validation procedure to reduce information leakage and optimistic performance estimates.

Predictive performance will be evaluated using cross-validated R-squared, mean absolute error, and root mean squared error. Higher R-squared values and lower MAE and RMSE values will indicate better predictive performance. The distribution and variability of performance estimates across cross-validation resamples will be examined.

Model-specific variable-importance analyses and, when applicable, variable-selection stability analyses will be used to identify variables that contribute most consistently to prediction. Variable importance will not be interpreted as evidence of a causal relationship. Because the present dataset permits internal validation only, all findings will be interpreted as exploratory and hypothesis-generating, and external validation will be required before any model can be considered for clinical use.