Intrauterine growth restriction (IUGR) is a complication that may arise during pregnancy. The studies show that the IUGR newborns may be with low IQ, dysfunction, or malnutrition; thus, it is essential to diagnose or predict the IUGR early during the pregnancy. To measure the size of the uterus, and then predict the size of the fetus is one of the existing methods for the diagnosis of IUGR. Also, ultrasound detection is widely used over the world. This study aims to discuss and analysisthe prediction of intrauterine growth restriction (IUGR) with the application of artificial intelligence techniques, to be exact, machine learning. Different from the existing methods, the machine learning for prediction IUGR based on the recordings of cardiotocography (CTG). This prospective cohort study selected 61 pregnant cCTG records (one trace for each patient) performed in IUGR pregnancies, and 61 healthy controls. For each record, it contains 13 features, which are: short-term variability (STV), long-term irregularity (LTI), Delta, approximate entropy (ApEn), Interval Index (II), spectral components as low frequency (LF), median frequency (MF), high frequency (HF), LF/(HF+MF) ratio, gestational age, Acceleration phase-rectified slope (APRS), deceleration phase-rectified slope (DPRS) and Lempel Ziv Complexity(LZC). As for the algorithms of machine learning for IUGR prediction, this study discussed the six types of classifiers: the logistic regression for binary classification, the k-nearest neighbor (KNN), the naïve Bayes, the support vector machine (SVM), the classification and regression tree (CART) and the random forest. The top three precise algorithms of machine learning for predicting IUGR based on cCTG records are the random forestrandom forestnd logistic regression, all of them are higher than 90%. Besides, the top three stable algorithms of machine learning on IUGR prediction based on cCTG records are the random forest, the CART, and the logistic regression. The random forest classification performs best on the IUGR prediction among these methods, with both highest accuracy and stability. The performance of CART is closed to the random forest; however, it usually over-fitting. While the logistic regression performs good, the performance of which is worse than the random forest to some degree.
La restrizione della crescita intrauterina (IUGR) è una complicazione che può insorgere durante la gravidanza. Gli studi dimostrano che i neonati IUGR potrebbero avere un QI basso, disfunzione o malnutrizione; quindi, è essenziale diagnosticare o prevedere la IUGR all'inizio della gravidanza. Misurare le dimensioni dell'utero e quindi prevedere la dimensione del feto è uno dei metodi esistenti per la diagnosi di IUGR. Inoltre, la diagnostica ad ultrasuoni è ampiamente utilizzata in tutto il mondo. Lo scopo di questo lavoro è di discutere e analizzare la previsione della restrizione della crescita intrauterina (IUGR) con l'applicazione di tecniche di intelligenza artificiale tramite machine learning. Diversamente dai metodi esistenti, il machine learning si basa sulle registrazioni di cardiotocografia (CTG) per la previsione dell’IUGR. Per questo studio di coorte prospettico vengono selezionate 61 registrazioni di cCTG gravide (una traccia per ciascun paziente) eseguite in gravidanze complicate da IUGR e 61 controlli sani. Ogni registrazione contiene 13 caratteristiche: variabilità a breve termine (STV), irregolarità a lungo termine (LTI), Delta, entropia approssimativa (ApEn), indice dell’intervallo (II), componenti spettrali come bassa frequenza (LF), frequenza mediana (MF), rapporto frequenza alta (HF), rapporto LF / (HF + MF), età gestazionale, pendenza rettificata in accelerazione (APRS), pendenza rettificata in fase di decelerazione (DPRS) e complessità Lempel Ziv (LZC). Per quanto riguarda gli algoritmi di machine learning per la previsione di IUGR, in questo studio vengono discussi sei tipi di classificatori: la regressione logistica per la classificazione binaria, il k-nearest neighbor (KNN), l'ingenuo Bayes, il supporto vector machine (SVM), la classificazione e albero di regressione (CART) e la foresta casuale. I migliori tre algoritmi per la previsione di IUGR sulla base di registrazioni cCTG, con i risultati superiori del 90%, sono la foresta casuale, il CART e la regressione logistica. Inoltre, i tre algoritmi più stabili di machine learning sono la foresta casuale, la CART e la regressione logistica. La classificazione foresta casuale risulta migliore per la previsione dell’IUGR, con la massima accuratezza e stabilità, tra tutti i metodi testati. La prestazione di CART è vicina alla foresta casuale; tuttavia risulta spesso over-fitting. La regressione logistica esegue bene, la cui performance è peggiore della foresta casuale in parte.
The application and analysis of machine learning for intrauterine growth restriction classification
MA, WENZHUO
2018/2019
Abstract
Intrauterine growth restriction (IUGR) is a complication that may arise during pregnancy. The studies show that the IUGR newborns may be with low IQ, dysfunction, or malnutrition; thus, it is essential to diagnose or predict the IUGR early during the pregnancy. To measure the size of the uterus, and then predict the size of the fetus is one of the existing methods for the diagnosis of IUGR. Also, ultrasound detection is widely used over the world. This study aims to discuss and analysisthe prediction of intrauterine growth restriction (IUGR) with the application of artificial intelligence techniques, to be exact, machine learning. Different from the existing methods, the machine learning for prediction IUGR based on the recordings of cardiotocography (CTG). This prospective cohort study selected 61 pregnant cCTG records (one trace for each patient) performed in IUGR pregnancies, and 61 healthy controls. For each record, it contains 13 features, which are: short-term variability (STV), long-term irregularity (LTI), Delta, approximate entropy (ApEn), Interval Index (II), spectral components as low frequency (LF), median frequency (MF), high frequency (HF), LF/(HF+MF) ratio, gestational age, Acceleration phase-rectified slope (APRS), deceleration phase-rectified slope (DPRS) and Lempel Ziv Complexity(LZC). As for the algorithms of machine learning for IUGR prediction, this study discussed the six types of classifiers: the logistic regression for binary classification, the k-nearest neighbor (KNN), the naïve Bayes, the support vector machine (SVM), the classification and regression tree (CART) and the random forest. The top three precise algorithms of machine learning for predicting IUGR based on cCTG records are the random forestrandom forestnd logistic regression, all of them are higher than 90%. Besides, the top three stable algorithms of machine learning on IUGR prediction based on cCTG records are the random forest, the CART, and the logistic regression. The random forest classification performs best on the IUGR prediction among these methods, with both highest accuracy and stability. The performance of CART is closed to the random forest; however, it usually over-fitting. While the logistic regression performs good, the performance of which is worse than the random forest to some degree.File | Dimensione | Formato | |
---|---|---|---|
Wenzhuo-Master Final Thesis.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Wenzhuo Ma-Master Final Thesis
Dimensione
1.27 MB
Formato
Adobe PDF
|
1.27 MB | 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/10589/149036