Nonlinear mixed-effects (NLME) models are an advanced tool widely used in the study of interactions between hormones and substrates, enabling the characterization of the pathophysiology of metabolic diseases. In NLME models, time-varying regressors are often used as input signals to the model and are commonly referred to as forcing functions. These regressors enable the description of dynamics that would otherwise be difficult to achieve without excessively increasing the complexity of the model. In current software for the identification of NLME models, the solution of the ordinary differential equations (ODE) that constitute the model requires the interpolation of these functions on a sufficiently dense time grid, typically using "zero-order hold" techniques. This approach is sub-optimal and leads to a consequent degradation of the estimates obtained, especially in the case of forcing functions affected by noise or exhibiting particularly fast dynamics. This study seeks to examine the process of identifying NLME models when forcing functions are present, and to suggest approaches for minimizing their effects. For this purpose, the Glucose Minimal Model (GMM) and the C-peptide Minimal Model (CMM) were considered as case studies, which use plasma insulin and glucose concentrations as forcing functions, respectively. The identification of minimal models was performed using Monolix 2024R1 and NONMEM 7.6.0, the two most widely used software for NLME analysis. Starting from a dataset comprising 204 subjects undergoing intravenous glucose tolerance testing, two synthetic datasets of 30 subjects each were generated, simulating the GMM and CMM respectively. After that, five additional datasets were constructed, in order to quantify algorithm performance in identifying the models on datasets with different characteristics in terms of signal to noise ratio and frequency sampling. The analysis of the results revealed that the method used to interpolate time-varying regressors can have a significant impact on the estimates. Therefore, three approaches were tested to improve the use of forcing functions by the NLME identification software. For both minimal models, the introduction of interpolated samples into the forcing function yielded the best parameter estimates in NONMEM while, in Monolix, the best results were obtained by interpolating the forcing function within the ODE solver. In addition, three potential limiting factors that could affect the evaluation of the model were investigated but ultimately discarded, as they did not produce any significant impact on the results. Future developments of this work may explore alternative approaches to improve the interpolation of the forcing function within the NLME identification software, with the goal of generating profiles with smoother and more realistic dynamics.
I modelli non lineari a effetti misti (NLME) sono uno strumento avanzato e ampiamente utilizzato nello studio delle interazioni tra ormoni e substrati, consentendo la caratterizzazione della fisiopatologia delle malattie metaboliche. Nei modelli NLME sono spesso utilizzati i regressori tempo varianti come segnali di input al modello, comunemente indicati come funzioni forzanti. Questi regressori consentono di descrivere dinamiche che altrimenti sarebbero difficili da ottenere senza aumentare eccessivamente la complessità del modello. Nei software attualmente utilizzati per l’identificazione dei modelli NLME, la soluzione delle equazioni differenziali ordinarie (ODE) che compongono il modello richiede l’interpolazione di queste funzioni su una griglia temporale sufficientemente densa, tipicamente mediante tecniche di "zero-order hold". Questo approccio è subottimo e comporta un deterioramento delle stime ottenute, soprattutto nel caso di funzioni forzanti affette da rumore o caratterizzate da dinamiche particolarmente rapide. Questo studio mira a esaminare il processo di identificazione dei modelli NLME in presenza di funzioni forzanti, e a suggerire approcci per minimizzare i loro effetti. A tal fine, sono stati considerati come casi di studio il modello minimo del glucosio (GMM) e il modello minimo del C-peptide (CMM), che utilizzano rispettivamente le concentrazioni plasmatiche di insulina e glucosio come funzioni forzanti. L’identificazione dei modelli minimi è stata effettuata utilizzando Monolix 2024R1 e NONMEM 7.6.0, i due software più utilizzati per l’analisi NLME. A partire da un dataset comprendente 204 soggetti sottoposti a test intravenoso di tolleranza al glucosio, sono stati generati due dataset sintetici di 30 soggetti ciascuno, simulando rispettivamente il GMM e il CMM. Successivamente, sono stati costruiti altri cinque dataset, al fine di quantificare le prestazioni degli algoritmi nell’identificazione dei modelli su dataset con differenti caratteristiche in termini di rapporto segnale-rumore e frequenza di campionamento. L’analisi dei risultati ha mostrato come il metodo di interpolazione dei regressori tempo varianti possa avere un impatto significativo sulle stime ottenute. Sono stati quindi testati tre approcci volti a migliorare l’uso delle funzioni forzanti da parte dei software per l’identificazione di modelli NLME. Per entrambi i modelli minimi, l’introduzione di campioni interpolati nella funzione forzante ha fornito le migliori stime dei parametri in NONMEM, mentre in Monolix i risultati migliori sono stati ottenuti interpolando la funzione forzante all’interno del risolutore ODE. Inoltre, sono stati analizzati tre potenziali fattori limitanti che avrebbero potuto influenzare la valutazione del modello, ma che sono stati successivamente scartati in quanto non hanno prodotto alcun impatto significativo sui risultati. Sviluppi futuri di questo lavoro potrebbero esplorare approcci alternativi volti a migliorare l’interpolazione della funzione forzante da parte dei software per l’identificazione di modelli NLME, con l’obiettivo di generare profili con dinamiche più regolari e realistiche.
Identificazione di modelli non lineari a effetti misti in presenza di regressori tempo varianti: il caso di studio dei modelli minimi di glucosio e C-peptide
TECCHIO, GIULIA
2024/2025
Abstract
Nonlinear mixed-effects (NLME) models are an advanced tool widely used in the study of interactions between hormones and substrates, enabling the characterization of the pathophysiology of metabolic diseases. In NLME models, time-varying regressors are often used as input signals to the model and are commonly referred to as forcing functions. These regressors enable the description of dynamics that would otherwise be difficult to achieve without excessively increasing the complexity of the model. In current software for the identification of NLME models, the solution of the ordinary differential equations (ODE) that constitute the model requires the interpolation of these functions on a sufficiently dense time grid, typically using "zero-order hold" techniques. This approach is sub-optimal and leads to a consequent degradation of the estimates obtained, especially in the case of forcing functions affected by noise or exhibiting particularly fast dynamics. This study seeks to examine the process of identifying NLME models when forcing functions are present, and to suggest approaches for minimizing their effects. For this purpose, the Glucose Minimal Model (GMM) and the C-peptide Minimal Model (CMM) were considered as case studies, which use plasma insulin and glucose concentrations as forcing functions, respectively. The identification of minimal models was performed using Monolix 2024R1 and NONMEM 7.6.0, the two most widely used software for NLME analysis. Starting from a dataset comprising 204 subjects undergoing intravenous glucose tolerance testing, two synthetic datasets of 30 subjects each were generated, simulating the GMM and CMM respectively. After that, five additional datasets were constructed, in order to quantify algorithm performance in identifying the models on datasets with different characteristics in terms of signal to noise ratio and frequency sampling. The analysis of the results revealed that the method used to interpolate time-varying regressors can have a significant impact on the estimates. Therefore, three approaches were tested to improve the use of forcing functions by the NLME identification software. For both minimal models, the introduction of interpolated samples into the forcing function yielded the best parameter estimates in NONMEM while, in Monolix, the best results were obtained by interpolating the forcing function within the ODE solver. In addition, three potential limiting factors that could affect the evaluation of the model were investigated but ultimately discarded, as they did not produce any significant impact on the results. Future developments of this work may explore alternative approaches to improve the interpolation of the forcing function within the NLME identification software, with the goal of generating profiles with smoother and more realistic dynamics.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tecchio_Giulia.pdf
accesso aperto
Dimensione
26.5 MB
Formato
Adobe PDF
|
26.5 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/98462