Background: Insulin pump failures (IPF) represent a major complication in the management of type 1 diabetes mellitus, as they can lead to rapid glycemic rise and diabetic ketoacidosis (DKA). Traditional mechatronic alarms are effective in detecting complete occlusions but often fail to promptly identify partial obstructions or reduced insulin delivery, resulting in clinically significant delays and frequent false positives. Recent evidence also reports an increased risk of DKA in patients treated with automated insulin delivery(AID). In this context, data-driven algorithms integrating glycemic signals, insulin delivery, and carbohydrate intake may represent a promising strategy for early malfunction detection and prevention of acute complications. Aims: The primary aim of this study was to validate, under controlled and real-life conditions, the ability of an unsupervised algorithm (Isolation Forest, iForest) to detect reduced or absent insulin infusion. Secondary aims included the characterization of glycemic and ketone dynamics during insulin suspension and the identification of clinical and metabolic factors associated with enhanced ketone response. Materials and Methods: This pilot study was divided into three phases: (1) 30-day free-living monitoring with collection of CGM data, insulin doses, and carbohydrate intake; (2) controlled suspension of insulin infusion for up to 6 hours, simulating pump failure, with measurements of glucose, insulin, and ketone levels; and (3) retrospective application of the iForest algorithm in a subgroup of patients using the Tandem t:slim X2 system with Control-IQ. Primary endpoints were sensitivity (SE) in detecting IPF and number of false positives per day (FP/day). Results: Twenty adults were enrolled (11 males, mean age 41.3 ± 13.3 years, BMI 26.5 ± 3.1 kg/m², diabetes duration 28.1 ± 10.3 years, HbA1c 6.8%); 19 completed the study, and 15 were analyzed with iForest. During insulin suspension, glucose increased by 33.6 mg/dL/h and ketones by 0.24 mmol/L/h. All participants reached 180 mg/dL in 2.4 h and 0.6 mmol/L of ketones in 2.7 h; 47% exceeded 1.5 mmol/L in 3.5 h. iForest showed 73.3% SE for induced IPF, 30% for manually labeled IPF, and overall SE 56%, with 0.12 FP/day (0.069 after excluding clinically acceptable alarms). Participants with maximal ketonemia ≤ 1.5 mmol/L had significantly higher body weight than those > 1.5 mmol/L (p = 0.025). The CGM/IOB rate ratio correlated significantly with ketone increase (p < 0.001). Conclusions: The iForest algorithm effectively identified IPF events with good sensitivity and a low false-positive rate. The relationship between glucose variation and residual active insulin (CGM/IOB rate) may serve as an early marker of ketone accumulation risk. Integrating clinical and metabolic parameters, including ketone monitoring, into future predictive algorithms could enhance the safety of automated insulin delivery systems and reduce DKA risk.
Background: Le occlusioni dei microinfusori (Insulin Pump Failures, IPF) rappresentano una complicanza rilevante nella gestione del diabete mellito di tipo 1, potendo causare rapido incremento glicemico e chetoacidosi diabetica (DKA). Gli allarmi meccatronici tradizionali rilevano efficacemente le occlusioni complete, ma non identificano sempre tempestivamente ostruzioni parziali, generando ritardi clinici e numerosi falsi positivi. Recenti evidenze segnalano inoltre un aumento del rischio di DKA nei pazienti trattati con sistemi ibridi ad ansa chiusa (AID). In tale contesto, l’impiego di algoritmi data-driven basati sull’analisi integrata di segnali glicemici, infusioni insuliniche e carboidrati dichiarati rappresenta una strategia promettente per il rilevamento precoce dei malfunzionamenti. Obiettivi: Validare, in condizioni controllate e di vita reale, la capacità di un algoritmo non supervisionato (Isolation Forest, iForest) di identificare ridotta o assente infusione di insulina. Obiettivi secondari erano la caratterizzazione dell’andamento glicemico e chetonemico durante la sospensione insulinica e l’identificazione dei fattori clinici e metabolici associati a maggiore risposta chetonemica. Materiali e metodi: Studio pilota articolato in tre fasi: (1) monitoraggio in vita reale per 30 giorni con raccolta di dati CGM, dosi insuliniche e carboidrati; (2) sospensione controllata dell’infusione insulinica fino a 6 ore, simulando un guasto del microinfusore, con prelievi seriati per glicemia, insulinemia e chetonemia; (3) applicazione retrospettiva dell’algoritmo iForest in pazienti utilizzatori di sistema Tandem t:slim X2 con Control-IQ. End-point primari: sensitività (SE) nel rilevare IPF e numero di falsi positivi per giorno (FP/day). Risultati: Arruolati 20 pazienti (11 maschi, età 41,3 ± 13,3 anni, BMI 26,5 ± 3,1 kg/m², durata diabete 28,1 ± 10,3 anni, HbA1c 6,8%). 19 hanno completato lo studio; 15 sono stati analizzati con iForest. Durante la sospensione insulinica, la glicemia è aumentata di 33,6 mg/dL/h e la chetonemia di 0,24 mmol/L/h. Tutti hanno raggiunto 180 mg/dL in 2,4 ore e 0,6 mmol/L di chetoni in 2,7 ore; il 47% ha superato 1,5 mmol/L in 3,5 ore. L’algoritmo ha mostrato SE del 73,3% negli IPF indotti, 30% negli IPF etichettati manualmente e SE complessiva del 56%, con 0,12 FP/day (0,069 escludendo gli allarmi clinicamente accettabili). I soggetti con chetonemia ≤1,5 mmol/L presentavano peso maggiore rispetto a quelli con valori >1,5 mmol/L (p = 0,025). Il rapporto rate CGM/IOB è risultato significativamente correlato alla crescita della chetonemia (p < 0,001). Conclusioni: L’algoritmo iForest ha individuato efficacemente gli IPF con buona SE e basso tasso di falsi positivi. Il rapporto tra variazione glicemica e insulina attiva residua potrebbe costituire un indicatore precoce di rischio chetonemico. L’integrazione di parametri clinici e metabolici nei futuri algoritmi predittivi potrà migliorare la sicurezza dei sistemi automatizzati di somministrazione di insulina, riducendo il rischio di DKA.
Valutazione clinica e validazione di un nuovo algoritmo per la gestione delle iperglicemie da occlusione nei sistemi di erogazione automatica dell'insulina (AID) in pazienti con diabete mellito di tipo 1
FURLANETTO, DAMIANO
2023/2024
Abstract
Background: Insulin pump failures (IPF) represent a major complication in the management of type 1 diabetes mellitus, as they can lead to rapid glycemic rise and diabetic ketoacidosis (DKA). Traditional mechatronic alarms are effective in detecting complete occlusions but often fail to promptly identify partial obstructions or reduced insulin delivery, resulting in clinically significant delays and frequent false positives. Recent evidence also reports an increased risk of DKA in patients treated with automated insulin delivery(AID). In this context, data-driven algorithms integrating glycemic signals, insulin delivery, and carbohydrate intake may represent a promising strategy for early malfunction detection and prevention of acute complications. Aims: The primary aim of this study was to validate, under controlled and real-life conditions, the ability of an unsupervised algorithm (Isolation Forest, iForest) to detect reduced or absent insulin infusion. Secondary aims included the characterization of glycemic and ketone dynamics during insulin suspension and the identification of clinical and metabolic factors associated with enhanced ketone response. Materials and Methods: This pilot study was divided into three phases: (1) 30-day free-living monitoring with collection of CGM data, insulin doses, and carbohydrate intake; (2) controlled suspension of insulin infusion for up to 6 hours, simulating pump failure, with measurements of glucose, insulin, and ketone levels; and (3) retrospective application of the iForest algorithm in a subgroup of patients using the Tandem t:slim X2 system with Control-IQ. Primary endpoints were sensitivity (SE) in detecting IPF and number of false positives per day (FP/day). Results: Twenty adults were enrolled (11 males, mean age 41.3 ± 13.3 years, BMI 26.5 ± 3.1 kg/m², diabetes duration 28.1 ± 10.3 years, HbA1c 6.8%); 19 completed the study, and 15 were analyzed with iForest. During insulin suspension, glucose increased by 33.6 mg/dL/h and ketones by 0.24 mmol/L/h. All participants reached 180 mg/dL in 2.4 h and 0.6 mmol/L of ketones in 2.7 h; 47% exceeded 1.5 mmol/L in 3.5 h. iForest showed 73.3% SE for induced IPF, 30% for manually labeled IPF, and overall SE 56%, with 0.12 FP/day (0.069 after excluding clinically acceptable alarms). Participants with maximal ketonemia ≤ 1.5 mmol/L had significantly higher body weight than those > 1.5 mmol/L (p = 0.025). The CGM/IOB rate ratio correlated significantly with ketone increase (p < 0.001). Conclusions: The iForest algorithm effectively identified IPF events with good sensitivity and a low false-positive rate. The relationship between glucose variation and residual active insulin (CGM/IOB rate) may serve as an early marker of ketone accumulation risk. Integrating clinical and metabolic parameters, including ketone monitoring, into future predictive algorithms could enhance the safety of automated insulin delivery systems and reduce DKA risk.| File | Dimensione | Formato | |
|---|---|---|---|
|
FURLANETTO DAMIANO TESI SPECIALIZZAZIONE.pdf
Accesso riservato
Dimensione
1.01 MB
Formato
Adobe PDF
|
1.01 MB | Adobe PDF |
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/20.500.12608/96673