Electronic Noses (E-Noses) are devices equipped with gas sensors, such as metal oxide (MOX) sensors, that detect volatile organic compounds (VOCs) and are used in environmental monitoring, food industry, and medical diagnostics. However, their widespread adoption is limited by sensor non-reproducibility, which necessitates individual calibration for each device and for a single device over time, limiting real world applications. For this reason, most E-Nose studies focus on short-term uses. Transferring calibration models from a "master" E-Nose to "slave" devices without full recalibration could save time and resources. This study proposes a method for transferring calibration models between nominally identical E-Noses, specifically for urine analysis in medical diagnostics. A PLS-DA model with VIP feature selection is developed to classify real urine, with added biomarkers to simulate pathologies. Synthetic calibrant mixtures are developed to mimic real urine responses in E-Noses and analyzed via MOS and PID sensors, aiming to stimulate such gas sensors as real samples do in conditions similar to the real application. These mixtures, reproducible and largely available, unlike real urine, are used as calibration standards to correct the responses of "slaves" through the analysis of only a few synthetic transfer samples on both devices. Direct Standardization (DS) is chosen as Calibration Transfer (CT) method, to build a standardized matrix to correct, time by time, the “slave” responses, enabling the use of a single model trained on the “master”. The tested transfer samples selection methods include Kennard-Stone, Extremes + Densest Cluster points and random selection. Four E-Noses (V1-V4) are used to study CT for four master-slave transfers: V1 as master with V2, V3, V4 as slaves (V1→V2, V1→V3, V1→V4) and V1+V2 as master (to study the impact of sensor duplication) with V3+V4 as slave (V1+V2→V3+V4). Without CT, prediction accuracies on the slaves are 37.14-54.93%. After applying DS, accuracies improve to 75.38-80%, comparable to the master’s one (78.95%). The similarity of synthetic mixtures to real urine is also evaluated in terms of sensory profile with a difference-from-control test and a descriptive test, identifying the best-mimicking recipes under that point of view.
I nasi elettronici (E-Nose) sono dispositivi costituiti da sensori di gas (es. MOX) sensibili ai composti organici volatili (VOC) e usati in ambiti ambientali, alimentari e medici. La non riproducibilità dei sensori richiede calibrazioni individuali di ciascun dispositivo nel tempo e per ogni nuovo dispositivo, limitando l'adozione degli E-Nose. Gran parte degli studi, infatti, si concentra su applicazioni a breve termine. L'uso di un modello addestrato per un E-Nose ("master") su altri dispositivi ("slave") senza una ricalibrazione completa ridurrebbe costi, tempo e risorse. Questo studio propone un metodo per trasferire modelli di calibrazione tra E-Nose nominalmente identici, con focus sull'analisi dell’urina per diagnosi mediche. Un modello PLS-DA con selezione features via VIP è addestrato per classificare urina reale con l'aggiunta di biomarcatori per simulare patologie. Miscele sintetiche sono sviluppate come standard di calibrazione per correggere le risposte degli "slave" analizzando pochi campioni sintetici, che a differenza dell’urina vera sono riproducibili e disponibili in larghe quantità. Queste miscele, analizzate tramite sensori MOX e PID, hanno l'obiettivo di stimolare i sensori come i campioni reali in condizioni simili all'applicazione reale. Con la Standardizzazione Diretta (DS), scelta per il Trasferimento di Calibrazione (CT), si crea una matrice standardizzata per correggere le risposte degli "slave" e usare un unico modello addestrato sul "master". I metodi di selezione dei campioni di trasferimento testati sono l'algoritmo di Kennard-Stone, una selezione di estremi + punti del cluster più denso e una selezione casuale. Quattro E-Nose (V1-V4) sono utilizzati per studiare il CT in quattro trasferimenti master-slave: V1 come master con V2, V3, V4 come slave (V1→V2, V1→V3, V1→V4) e V1+V2 come master (per studiare l'impatto della duplicazione dei sensori) con V3+V4 come slave (V1+V2→V3+V4). Senza CT, le accuratezze di predizione sugli slave sono del 37,14-54,93%. Dopo l'applicazione della DS, le accuratezze migliorano al 75,38-80%, comparabili a quella del master (78,95%). Le miscele sintetiche sono anche valutate per somiglianza sensoriale all’urina, identificando le ricette che meglio ne mimano il profilo olfattivo con un test dei descrittori e un test di differenza dal controllo.
Addressing the issue of gas sensor reproducibility: development of a calibration transfer methodology for urine headspace analysis
Cassinerio, Michela
2023/2024
Abstract
Electronic Noses (E-Noses) are devices equipped with gas sensors, such as metal oxide (MOX) sensors, that detect volatile organic compounds (VOCs) and are used in environmental monitoring, food industry, and medical diagnostics. However, their widespread adoption is limited by sensor non-reproducibility, which necessitates individual calibration for each device and for a single device over time, limiting real world applications. For this reason, most E-Nose studies focus on short-term uses. Transferring calibration models from a "master" E-Nose to "slave" devices without full recalibration could save time and resources. This study proposes a method for transferring calibration models between nominally identical E-Noses, specifically for urine analysis in medical diagnostics. A PLS-DA model with VIP feature selection is developed to classify real urine, with added biomarkers to simulate pathologies. Synthetic calibrant mixtures are developed to mimic real urine responses in E-Noses and analyzed via MOS and PID sensors, aiming to stimulate such gas sensors as real samples do in conditions similar to the real application. These mixtures, reproducible and largely available, unlike real urine, are used as calibration standards to correct the responses of "slaves" through the analysis of only a few synthetic transfer samples on both devices. Direct Standardization (DS) is chosen as Calibration Transfer (CT) method, to build a standardized matrix to correct, time by time, the “slave” responses, enabling the use of a single model trained on the “master”. The tested transfer samples selection methods include Kennard-Stone, Extremes + Densest Cluster points and random selection. Four E-Noses (V1-V4) are used to study CT for four master-slave transfers: V1 as master with V2, V3, V4 as slaves (V1→V2, V1→V3, V1→V4) and V1+V2 as master (to study the impact of sensor duplication) with V3+V4 as slave (V1+V2→V3+V4). Without CT, prediction accuracies on the slaves are 37.14-54.93%. After applying DS, accuracies improve to 75.38-80%, comparable to the master’s one (78.95%). The similarity of synthetic mixtures to real urine is also evaluated in terms of sensory profile with a difference-from-control test and a descriptive test, identifying the best-mimicking recipes under that point of view.File | Dimensione | Formato | |
---|---|---|---|
2025_04_Cassinerio_Tesi.pdf
accessibile in internet solo dagli utenti autorizzati
Dimensione
5.58 MB
Formato
Adobe PDF
|
5.58 MB | Adobe PDF | Visualizza/Apri |
2025_04_Cassinerio_Executive Summary.pdf
accessibile in internet per tutti
Dimensione
885.4 kB
Formato
Adobe PDF
|
885.4 kB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/236262