Model predikcije uspešnosti indukcije porođaja zasnovan na kliničkim i ultrazvučnim parametrima trudnice

Abstract

Uvod: Postoji trend porasta incidence indukcije porođaja i rezultati pokazuju da je oko 20% porođaja u svetu indukovano. Ishod indukcije porođaja nije izvestan, a sama indukcija porođaja nosi određene rizike. Neophodni su modeli za predviđanje ishoda indukcije porođaja, zasnovani na individualnim karakteristikama pacijentkinja i koji se sa velikom pouzdanošću mogu koristiti prilikom odluke o načinu porođaja. Izdvajanje trudnica koje su u visokom riziku od neuspele indukcije porođaja dovelo bi do redukcije maternalnog i neonatalnog morbiditeta i mortaliteta. Cilj: Glavni cilj studije bio je razvoj modela zasnovanog na algoritmima mašinskog učenja za predviđanje uspeha indukcije porođaja. Pre toga procenjeni su klinički i ultrazvučni parametri koji utiču na uspešnost indukcije porođaja, a zatim je razvijen i novi ultrazvučni sistem bodovanja zrelosti grlića materice (USS). Materijal i metode: Prospektivna randomizirana studija, obuhvatila je 226 trudnica, kojima je indukovan porođaj. Analizirani su klinički i ultrazvučni parametri trudnice pre indukcije porođaja. Indukcija porođaja podrazumevala je aplikaciju Prepidil® gela (Burnet skor < 6) ili intravenska administracija oksitocina (Burnet skor ≥ 6). Nakon evaluacije ultrazvučnih paramatera, kreirali smo ultrazvučni skoring sistem. Zatim je vršena procena USS u odnosu na Burnet skor i kliničke parametre. Na kraju je razvijen sveobuhvatni model koji koristi algoritme mašinskog učenja za predviđanje uspeha indukcije porođaja. Tehnika KFold Cross unakrsne validacije korišćena je za internu validaciju modela. Rezultati: Većina trudnica u istraživanju porođena je vaginalnim putem −158 (69,9%), dok je 68 trudnica (30,1%) porođeno carskim rezom. Najčešća indikacija za indukciju porođaja bila je postterminska trudnoća – 124 (54,87%), a u odnosu na metod indukcije porođaja dominirala je kombinacija amniotomije uz primenu oksitocina – 159 (70,35%). U pogledu kliničkih parametara, na ishod indukcije porođaja najveći uticaj imaju paritet (p<0,05), BMI (p<0,05) i Burnet skor (p<0,001). Svi ispitivani ultrazvučni parametri pokazali su statističku značajnost u odnosu na uspeh indukcije porođaja (p<0,05), osim zadnjeg cervikalnog ugla (p=0,861) i ultrazvučno procenjene telesne mase ploda (p=0,766). Parametri uključeni u USS bili su komparabilni sa parametrima Burnet skora. USS sadrži pet ultrazvučnih parametara: dužina grlića materice i tunelizacija, položaj potiljka ploda, zadnji cervikalni ugao i visina fetalne glavice. U svim cut off vrednostima USS je pokazao veću senzitivnost, a Burnet skor veću specifičnost. U odnosu na predviđanje vremena od početka indukcije porođaja do samog porođaja, USS je pokazao bolje rezultate u odnosu na Burnet skor (0,069: 0,482). Vrednost površine ispod ROC krive bila je viša za USS (AUC 0,734) u odnosu na Burnet skor (AUC 0,66). Najbolji rezultati dobijeni su kombinacijom Burnet skor+USS+klinički parametri (AUC 0,920). Radi jednostavnije primene u svakodnevnom radu i eksterne validacije modela, kreirani su stablo odluke i aplikativni softver predikcije uspeha indukcije porođaja. Zaključak: Ultrazvučni skoring sistem je jednostavan, lak za primenu, ima visoku senzitivnost i specifičnost što ga čini lako primenljivim u svakodnevnom kliničkom radu. Kreirani model, koji obuhvata kliničke parametre, ultrazvučne parametre, Burnet skor i koristi algoritme mašinskog učenja daje bolje rezultate od modela koji koriste druge parametre. Ipak, potrebna su dalja istraživanja kako bi se prevazišla ograničenja ove studije i potrvdila pouzdanost modela.

Introduction: There is a growing trend toward induction of labor and approximately 20% of labors are induced. The induction of labor is not always successful and carries risks. Models for predicting the outcome of induction of labor are necessary, based on the individual characteristics of patients and which can be used with high reliability when deciding on the method of delivery. Isolating pregnant women who are at high risk of failed induction of labor would lead to a reduction in maternal and neonatal morbidity and mortality. Objective: The main aim of this study was to develop a machine-learning-based model for predicting the success of labor induction. To that end, the clinical and ultrasound parameters that affect the successfulness of labor induction were assessed, and a new ultrasound scoring system (USS) was developed and then assessed. Study design: This prospective randomised study from a single centеr included 226 term women who underwent induction of labor. First, a wide range of clinical and ultrasound pre-induction parameters were recorded. The induction was initiated by endocervical administration of Prepidil® gel (for Burnett score <6) or with intravenous Oxytocin (for Burnett score ≥ 6). After evaluation of ultrasound parameters, we created ultrasound scoring system. Finally, a comprehensive model using machine learning algorithms for predicting the success of the induction of labor was developed. A Stratified KFold cross-validation technique was used to check the developed model internally and to prevent overfitting. Results: The majority of pregnant women were delivered vaginally - 158 (69.9%), while 68 (30.1%) pregnant women were delivered by caesarean section. The most common indication for labor induction was postterm pregnancy – 124 (54.87%), and the most common method of labor induction was a combination of amniotomy and oxytocin - 159 (70.35%). In terms of clinical parameters, this study found that induction of labor correlates with parity, BMI (both at p<0.05) and the Burnett score (p<0.001). All ultrasound parameters were statistically significant (p<0.05) except for the posterior cervical angle (p=0.861) and ultrasound-estimated fetal weight (p=0.766). The parameters included in the USS were comparable to the Burnett score parameters. Therefore, the new USS encompasses the following five ultrasound parameters: cervical length, the size of the posterior cervical angle, the funneling (if present), the position of the fetal occiput and the distance from the head of the fetus to the outer cervical os. In all cut off values, USS showed higher sensitivity, and Burnet score showed higher specificity. In relation to predicting the time from the start of labor induction to delivery itself, USS showed better results compared to the Burnett score (0.482 vs 0.069). The value of the area under the ROC curve, was higher for USS (AUC 0.734) compared to Burnett score (AUC 0.66). The best results were obtained with the combination of Burnett score+USS+clinical parameters (AUC 0.920). A decision tree and application software for predicting the success of labor induction were created, in order to enable the simplest possible use in daily work as well as external validation of the model. Conclusion: The ultrasound scoring system is simple, easy to use, has high sensitivity and specificity, which makes it easy to apply in daily clinical work. The findings imply that the model developed in this study, which takes into account clinical parameters, the ultrasound parameters and the Burnett score and uses machine learning algorithms, yields better results than models using other parameters. Nevertheless, further research is needed to overcome the limitations of the present study and confirm the reliability of the model.