(181) Estimating personalized lung function trends from routine clinical care to facilitate analysis of daily data capture through remote technologies

Background: Automated data capture through remote monitoring via apps and/or wearable devices has enabled researchers to capture daily biometric data, allowing for the investigation of personalized trends and novel predictors of clinical outcomes. An example is the remote capture of physical activity levels using Fitbit activity trackers. However, it is challenging to interpret associations between daily predictors and outcomes that are measured infrequently, for example FEV_1. FEV₁ is typically measured quarterly at clinical encounters, which may not match with data collected from daily remote monitoring. As a result, daily data must be collapsed into a small number of categories to match with the time points of FEV₁ measurements, ignoring the granular detail. The aim of this study was to investigate the optimal parameters for using flexible polynomial regression to estimate personalized trends in FEV₁ over time in children and adolescents with CF to facilitate detailed analysis of frequently collected novel outcomes.

Methods: Anonymized longitudinal laboratory spirometry measurements from children with CF were obtained from Great Ormond Street Hospital in London UK. FEV₁ was converted to %predicted using the GLI equations. FEV₁ measurements >1y apart and individuals with < 2 FEV₁ measurements were excluded. A function was created in R software to calculate polynomial regression models from a 1^st up to a 5^th order for predicting FEV₁ over time for every individual; the model with the lowest AIC for each individual was selected as the best fit. The optimal data parameters for the function to estimate personalized trends in FEV₁ were investigated.

Results: There were 9,312 FEV₁ measurements from 321 people 3-18y between 2011-2021. After exclusions there were 8,924 measurements from 302 individuals. There was an average of 30 values per person (range: 2-150), and a mean of 53 days between consecutive measures (range: 1-365). Most repeated measures were within 6 months (98%). For individuals with fewer values (n< 5), a linear model was deemed best since the polynomial predictions between measurements deviated further from observed data than for individuals with a higher number of values (n≥5). For individuals with 5-15 values, the predictions were more robust when the time between measurements was < 6 months, whereas time did not appear to influence polynomials for individuals with ≥15 values. As a result of the findings, the final polynomial function is conditional on both the number of FEV₁ values and the time interval between measurements.

Conclusions: The function successfully estimated personalized FEV₁ trends (Figure). It has potential future use in studies that use remote daily data capture for predictors but rely on routine, yet infrequent, clinical encounters for collecting data on outcomes.

Acknowledgements:

References: