Newborn skin maturity models for gestational age prediction

A multicenter clinical trial evaluated the accuracy of a novel device to detect preterm newborns. A portable multiband reectance photometric device assessed 781 newborns’ skin maturity and used machine learning models to predict reference gestational age, adjusting it to birth weight and antenatal corticosteroid therapy exposure. The day difference between the reference and the test had a median of -1.4 (IQR: -2.1). Using established methods such as comparator ultrasound and last menstrual period (LMP), the medians were 0 (IQR: 4) and 0.01 (IQR: 4), respectively. For prematurity discrimination, the area under the receiver operating characteristic curve (AUROC) was 0.986 (95% CI: 0.977 to 0.994). In newborns with absent or unreliable LMP, the intent-to-discriminate analysis showed that the test generated correct classications 95.8% of the time. The assessment of the newborn's skin maturity adjusted by learning models promises accurate pregnancy dating at birth without the use of antenatal ultrasound reference.


Introduction
Public policies and best practices advise pregnant women to plan pregnancy with early access to prenatal care for pregnancies to be safely monitored until birth. 1 However, many barriers to covering all pregnancies and births with due care remain unsurpassed, particularly in scenarios without well-equipped facilities. 2,3 Adverse maternal and neonatal outcomes affect newborns unevenly according to the birth scenario and gestational age. 4 First-day-of-life mortality is 30 times higher in low-and medium-income countries than in high-income countries. 5 Preterm neonates are more vulnerable to death or survival with neurological sequelae. The need to recognize early risks at birth is faced with reduced early access to prenatal care and high-cost technologies for pregnancy dating, such as obstetric echography in resourceconstrained settings. 6,7 Early obstetric ultrasound currently offers the best gestational age method. 8 However, access to high-cost equipment, poor training, lack of skills of health professionals, or late prenatal care limit pregnancy dating and, consequently, detection of prematurity. 7,9 Improving preterm birth outcomes demands accurate gestational age assessment to direct opportune decisions regarding neonatal care. 10,11 Approaches to enhance the reliability of pregnancy dating through more accurate and accessible technologies can improve pregnancy outcomes and neonatal survival. 8,12 Health technology development is critical for supporting healthcare systems. Medical devices and digital health technologies have brought innovative solutions with the potential to save lives, 13 mitigating gaps of quality among disparate healthcare scenarios. 14 Furthermore, digital health technologies emerge with the potential to impact the equality of healthcare, creating new landscapes of opportunities such as applied data science to improve prediction models. 15,16 Currently, computer science has advanced with improvements to medical practice, detecting patterns by processing datasets through layered mathematical models, 17 fostering skills and competencies of professionals to support the best healthcare decisions. 18 The new test explored in this study is an innovative approach used to estimate gestational age based on the photobiological properties of the newborn's skin and using learning predictive models enhanced with clinical variables. 19 Usable as a medical device, we developed this technology to easily assist health professionals in making better decisions regarding newborn care whenever the pregnancy dating is unknown or doubtful. This study aims to validate the photobiological model of skin maturity adjusted to clinical data to promptly detect gestational age and determine its accuracy in detecting prematurity. We tested the hypothesis of equivalence between the gestational age measured by this new test, the pregnancy dating comparators calculated using ultrasound exams, and the last menstrual period (LMP).

Results
Of the 791 potential eligible newborns, two were under Rh alloimmunization during pregnancy, which was considered an exclusion criterion (Fig. 1). Among the 789 newborns with their skin assessed with the medical device, eight had no reference standard to allow the dependent variable; four had no mandatory embryo measurement; three had no second ultrasound comparator; and one had an unsolved digit date error. All 781 newborns who met the eligibility criteria of the clinical trial were included in the analysis. According to the set of models, two predictive algorithms provided gestational age values, as detailed in the Gestational age prediction by the new test subsection.
Seven hundred four women gave birth to 781 newborns. Despite early access to prenatal care (median 9, IQR 4 weeks of gestation), only 296 (48.2%) women ful lled the criteria for valid LMP, among 614 who were able to provide such a date (Table 1). A number of the pregnant women (n= 215, 30.7%) received antenatal corticosteroid therapy for fetal maturation (ACTFM), following local protocols, corresponding to 273 / 777 (35.1%) newborns. According to reference gestational age at birth, 415 (53.1%) newborns were term. Among 366 (46.9%) preterm newborns, 235 (30.1%) had a gestational age at birth less than 37 to 32 weeks, and 131 (16.8%) had a gestational age less than 32 to 24 weeks, achieving approximately the expected arranged proportion for the sampling. 20 The frequency of fetal abnormal growth classi cation at birth was 116 (14.9%) who were small for gestational age and 59 (7.6%) who were large for gestational age.
Gestational age using established methods of pregnancy dating The distribution of gestational age calculated according to the different bases, reference (R), a second ultrasound exam after 13 weeks and 6 days of gestation and before 22 weeks (C1), and LMP (C2) corroborated the existing differences among established methods of antenatal dating gured in the overlapped histogram (Fig. 2). There were 27 (3.5%) posterm births when using comparator C2, 2 (0.3%) when using comparator C1, and 1 (0.1%) when using standard R for gestational age dating. Concerning the data quality of medical history reference used in gestational age estimation, the analysis of day preference con rmed bias for the LMP reference, the analysis of which is in Supplementary Note 1.
Correlation between the skin re ectance of the newborn and the reference gestational age Newborn skin re ection acquisition had a positive high correlation with the reference standard gestational age, Pearson correlation coe cient = 0.79, R 2 = 0.62, p < 0.001 (Fig. 3).

Gestational age prediction by the new test
We considered the best parsimonious models according to available variables in the birth scenario. When the birth weight was known and ACTFM exposure information was not, Model 1 tted gestational age with a linear-ridge algorithm. In cases when the birth weight and ACTFM information were known, Model 2 tted gestational age with an XGBoost algorithm. The set of models adjusted with machine learning models is presented in Supplementary Table in Note 2.
The boxplots in Fig. 4a present gestational ages calculated by applying the predictive algorithms: Rgestational age and the comparators C1-and C2-gestational age, highlighting preterm birth rates. Data concerning the gestational age distribution of frequency according to the methods are in Table 2. The agreement between Model 1, Model 2, and the R-gestational age was high, as well when analyzing predicted gestational ages against the comparators C1-and C2-gestational age. The Bland-Altman 95% limits were -28.6 to 28.6 days for Model 1 and -19.6 to 19.6 days for Model 2 (gestational age), Table 3.
With a focus on the day difference between methods of determining R-gestational age, Fig. 4b displays box plots for each method of estimation. The day difference had a median of -1.4 (IQR: -2.1) for the Model 2-gestational age, and a median of 0.21 (IQR: 14) days for the Model 1 gestational age ( Fig. 4b and Table 2). Established methods of pregnancy dating (C1 and C2) had median differences of 0 (IQR: 4) and -0.01 (IQR: 4), respectively (Table 2). Moreover, the proportion of preterm newborns correctly detected at birth within a one-week error is presented as the value within the interval between hatched lines (Fig.  4b). We noticed that Model 2-gestational age achieved 98.7% agreement with the reference pregnancy dating within one-week error, and Model 1-gestational age had 53.5%. Thus, the ACTFM exposure information added advantages to the skin re ectance and birth-weight-based Model 2.
Accuracy of the new test to discriminate preterm newborns Skin re ectance itself had good accuracy in discriminating preterm term newborns, with AUROCs of 0.956 (95% CI: 0.942, 0.970) at 37 weeks, 0.937 (95% CI: 0.916, 0.958) at 32 weeks, and 0.959 (95% CI: 0.936, 0.981) at 28 weeks (cutoff) (Supplementary Fig. 1). The overall accuracy of preterm newborn classi cation is presented in Fig. 5 according to the predictive Model 1-and Model 2-gestational age considering different prematurity cutoff points. Regarding the 95% con dence interval (CI) analysis for AUROC, Model 2 gestational age had higher performance in discriminating preterm against term newborns at 37 weeks than Model 1 gestational age (Fig. 5). For 32-and 28-week cutoffs, Model 1-and Model 2-gestational age had similar AUROCs, considering an overlapping of 95% CIs. A comprehensive analysis of prediction accuracy for preterm newborns by the methods of gestational age estimation for different prematurity cutoffs is shown in Table 3.
Intent to preterm newborn discrimination by the new test Birth care settings where the new technology is to be applied deserve an intent to preterm newborn discriminant analysis, simulating the existence of baseline references for the gestational age calculation. Therefore, we considered the lack of LMP recall or unreliable information as scenario one, corresponding to 57.7% of the sample ( Table 4). The lack of a reliable LMP resulted in low accuracy (69.6%) of discrimination with imprecise dates. Nevertheless, 86.9% of the newborns were correctly classi ed as preterm or term using Model 1-gestational age. Using gestational age based on Model 2, 95.8% of the newborns were correctly classi ed as preterm or term, with a 0.5% false-positive rate. Great accuracy using any available method for gestational age estimation was observed in scenario two, where the LMP was reliable.

Safety of the device
There were no reports of unexpected medical events, unintended illness or injury, or unfortunate clinical signs in subjects, users, or others related to the investigational product or otherwise.

Discussion
A reliable antenatal age is a prerequisite for preterm newborn classi cation in birth care settings and constitutes the rst step to delivering the necessary care, considering the risks of prematurity. 11 Term newborns, allied with good tone, breathing or crying, are essential elements to determine steps of newborn resuscitation. 21 Although that statement in itself seems very simple, the reality is far from it.
Without the certainty of the day in the female cycle on which conception occurred, ultrasound measurement of the crown-rump length is a standard consensual reference for pregnancy dating. 8 This dependence on early echographic scans has deprived many pregnant women and their babies of reliable gestational age. 10 Such a technological gap causes even more disparities than the difference between childbirth scenarios in fully equipped facilities and those ill equipped with scarce technology. Moreover, it can impair the correct classi cation of infants as premature or growth-restricted. 22,23 The main contribution of this clinical trial is to validate a new approach for gestational age estimation independent of fetal ultrasound measures by demonstrating highly accurate outcomes. Based on two pieces of databirth weight and ACTFM exposure-and the use of a frugal medical device to assess skin maturity and process algorithms, 359 of 366 preterm neonates with less than 37 weeks of gestation were detected, with 96% being correctly classi ed.
In this combined study covering enhancing and validating prediction models, we believe the application of k-fold cross-validation with the use of machine learning algorithms provided accurate predictions. 24 While large data samples are unavailable, the process of training and testing are able to estimate the performance of algorithms until we have nished other ongoing clinical trials for external validation. 25 Furthermore, the quanti cation of uncertainty intervals regarding the predicted gestational age (calculated in days) and comparisons with established references allowed the simulation of real scenarios for application. Furthermore, the con dence intervals accompanying AUROC's accuracy contributed to revealing the forecast's limits as to discriminating term from preterm newborns at different cutoff points, with clinical relevance. Such strengths are critical to assure the potential value of the new test in facing the challenges of postnatal prematurity identi cation. 26 A new type of data science algorithm has thus emerged with the aim of qualifying pregnancy dating. High-performance reports using learning models based on antenatal ultrasound predictors 27 contradistinguished meager outcomes from those using other morphometric postnatal predictors. 28 Moreover, valuable algorithms with postnatal combinations on maturity scores of newborns are promising, demanding special skills to apply them. 29 Underquali ed birth attendants represent a challenge in developing countries, further limiting the use of existing birth care solutions. 30 One advantage of our new device is the skin assessment automation that alerts measurement error caused by movement of the newborn or examiner. Previous reports have detailed the human skin's light-skin interaction and optical properties that bene t this technology. 19,31 The predictive algorithms used information that health professionals could quickly obtain in childbirth settings-the birth weight and the ACTFM exposure-and which could add value to the physical data of skin maturity. Moreover, the device is capable of providing a gestational age to overcome extreme situations without antenatal records to obtain ACTFM exposure information. However, the algorithm of Model 1 had more comprehensive Bland-Altman 95% limits of 28.6 days compared to the 19.6 days of Model 2 with the full three variables. Working with more exible forecasts within seven days of error range, 98.7% of newborns had a valuable prediction with Model 2. Considering the simulated scenario with absent or unreliable LMP and lack of ACTFM exposure data-57.7% (n=451) of the sample-the algorithm of Model 1 for gestational age estimation detected 180 of 199 (90.5%) preterm newborns, even with low speci city (48.0%). This result expands the context of use for this medical device since the gestational age based on memory recall of the LMP missed 69 out of 199 preterm newborns, expressing a lower sensitivity (65.8%) when we applied the intent-to-discriminate analysis.
The choice and analysis of two algorithms for gestational age prediction (Model 1 and Model 2) depended on uncertainties regarding the effect of corticosteroids on forecasts. Antenatal corticosteroids to improve newborn outcomes are a practical, evidence-based intervention recommended for women at risk of preterm birth. 32 However, even with the acceleration of lung maturity, the effect of the drug occurs in other organs. The early fetal presence of receptors of corticosteroid hormone receptors in skin epithelial cells indicates that glucocorticoids may play an important role in the differentiation and development of human skin. 33 However, clinical evidence of the effect of ACTFM exposure on skin maturity is weak, and the topic remains unsubstantiated. 34 Thus, until proven otherwise, we interpreted that the importance of ACTFM exposure data to better adjust the gestational age modeling is related to an effect on skin maturity. Even so, we cannot deny that antenatal exposure to corticotherapy is more common in premature infants-72.3% of preterm newborns in this sample. In this respect, this predictor variable could imply a bias favoring preterm newborn detection. The aforementioned ongoing study for external validation of the algorithms could further elucidate this issue because the enrollment process of newborns introduced the Mozambican birth scenario, where unfortunately ACTFM is not guaranteed for every woman at risk of preterm birth. 25 Birth weight is a known estimator of risks to newborns. As part of primary routines in childbirth settings, this information has practical applicability, even in facilities with scarce high-cost technologies. 1 Meanwhile, predicting preterm birth based on birth weight when lacking a gold standard is an imperfect solution. 9 Additionally, the LMP reference and the postnatal scores of newborn maturity have demonstrated low accuracy in determining gestational age and identifying prematurity. 35 Later prenatal care and unquali ed date recollection justify efforts to enhance the reliability of pregnancy dating through more accurate and accessible technologies, seeking to improve pregnancy outcomes and neonatal survival. 10 In our study, qualifying the LMP at birth with questions about memory of date, menstrual cycles, and checking antenatal clinical documents at birth provided a gestational age able to identify 160/167 (95.8%) preterm newborns. Current approaches to calculating gestational age are sensitive to data quality, resulting in misplaced classi cation of prematurity. 9 The present study was committed to representing a real scenario in terms of data quality, as stated in the research protocol, with data collection and curation to assure the best reference and comparators for the analysis. Before opening the blinding of the trial, a consistency process confronted data entries with digital images of clinical documents taken during the enrollment. Furthermore, dedicated software was developed exclusively for the clinical trial, considering the quality of the variables and their constraints. Part of the enrollment occurred during the COVID-19 pandemic, causing a minimal amount of missing data, such as ACMF information (4/781 newborns).
Regarding the generalizability of outcomes, this multicenter trial gathered perinatal centers from the northern, central, southwestern, and southern regions of Brazil. This collaborative evaluation contributed to sampling a mixed population of newborns with high miscegenation and involved 15 examiners who attended good clinical practice training. The intraobserver error and interobserver error of measures were slow, corroborating previous results. 19 The number of preterm newborns was enough to analyze subcategories of prematurity as extreme preterm (n=42); however, the overall rate of preterm newborns was 46.9%, values observed in referral facilities and not in the general population of Brazilian newborns. 36 The number of neonatal deaths during 72 h of follow-up was 14, with 12 deaths occurring in newborns with gestational age below 28 weeks due to extreme prematurity complications. We expect to target worse childbirth scenarios for this technology implementation. 30 In addition, the safety of this device is similar to other optical technologies already used in neonatal care, such as transcutaneous bilirubinometer and pulse oximetry. 21,37 With regard to limitations, the adoption of the new test deserves caution. Impaired fetal growth and largefor-gestational age newborn in uence were not included in the analysis. Nevertheless, photometer-based technology for skin maturity assessment has a basis on skin transparency in part associated with epidermal thickness. 19 In a previous analysis of high-frequency ultrasound of newborn skin, we reported that the epidermal layer had no signi cant in uence on the fetal growth pattern when associated with gestational age. 38 Furthermore, the epidermal thickness of newborns was itself a predictor of gestational age when analyzing postmortem histological skin slices from the sole. 39 For the future, comparisons are expected based on postnatal approaches for gestational age estimation, such as scores of maturity and foot length or image combinations. 28 Identifying preterm newborns is the rst step to attending to their needs. The global rate of neonatal mortality corresponds to 6,700 neonatal daily deaths, mostly from preventable or treatable conditions in scenarios of healthcare scarcity. 40 We hope that strengthening the data sources of healthcare facilities with a reliable gestational age can help in identifying vulnerable newborns in situations with the absence or lack of such information.

Study design and Participants
A multicenter prospective clinical trial intention-to-diagnosis study by single group, single blinding, and single arm with a reference standard. This article adheres to the Transparent Reporting of a Multivariable Prediction Model for Individual Prediction or Diagnosis (TRIPOD) for completeness and clearness. 41 To assess the risk of bias and applicability, the development and validation methods followed guidance from the Prediction Model Risk of Bias Assessment Tool (PROBAST). 42 The clinical trial protocol is disclosed in the Brazilian Clinical Trials Registry (ReBec) RBR-3f5bm5.
This report examined primary and secondary outcomes of data concerning prematurity prediction and safety. Other secondary outcomes related to lung maturity prediction are under analysis for further participants' timeline, and the procedures followed the clinical protocol. 20 In short, we assessed the skin maturity of live newborns with at least 24 weeks or more of gestational age. All of them had their gestational age estimated by clinical parameters, with the embryo measurement mandatorily assessed by ultrasound exam at <14 weeks of gestation. Anhydramnios, hydrops, congenital skin diseases, or chorioamnionitis were the exclusion criteria due to their potential to modify the skin structure.

Procedures
The coordinator center trained 15 examiners following good clinical practice as set forth by the Brazilian regulatory health agency's recommendations. 43 In loco reliability of examiners was assessed during the certi cation visit of a senior researcher in the collaborator centers (Supplementary Note 3). Standard operating procedures were mandatory to advise enrollment. Data collection was conducted with the use of a smartphone and skin assessment with the medical device under validation. 43,45 An automated algorithm in the data collection system, 46 blinded to the examiner, calculated the Rgestational age. For this, a set of rules adjusted the due date using data entries collected from the ultrasound report or prenatal care book/document, providing the best reference. 8 Clinical information was collected in structured questionnaires using a software program dedicated to this project. The framework of clinical variables and skin acquisitions is available in Supplementary Note 4. For data curation, the investigator's data entries were confronted with information from photographed digital images of clinical documents. In the case of multiple birth gestations with different ultrasonographic crown-rump-length values, the average of each embryo/fetal value was considered. A double-check system-paper-based and electronic-allowed the veri cation of the reliability and validity of clinical data as well as skin thickness acquisition. In addition, the data quality of antenatal pregnancy dating was evaluated by comparing the frequency of days in dates, as they should be random with no preference for digits. For this, in cases of multiple gestations, we retained only the rst twin information for day-digit evaluation (Supplementary Note 1).

The Intervention
The intervention is a new device that processes the backscattered signal acquired from the skin of the newborn's sole with clinical variables to predict gestational age. Its development included steps from the workbench to the clinical experimentation already described. 19 Likewise, we previously analyzed the best body position to assess the skin re ectance for pregnancy dating, as well as environmental in uences such as humidity, temperature, ambient light, and the newborn's skin color. 19,31,38 Regarding the characteristics of the components, wavelengths from 400 nm to 1200 nm of the light emitter categorized the safety level of this medical device as Class II (noninvasive and medium risk) according to the regulatory agency in Brazil. 44 When the light emitting sensor touches the skin over the sole for a few seconds, it triggers 10 automated measurements. The device emits alerts regarding measurement errors caused by involuntary movement of the newborn or examiner under a set of known constraints of the skin re ectance of a newborn. 19 Such an event requires a new attempt. The device output was blinded to the examiners.
The skin assessment occurred with the newborn inside incubators, incubators-radiant warmer, warming pad-bassinet, standard crib, or in the mother's lap to ensure minimum manipulation and avoid unbalancing the clinical conditions. The sensor touched the sole three times, following complete disinfection with alcohol. Fourteen minimum viable products were produced for the study (Fig. 6). At the beginning and the end of the clinical trial, the irradiance emitted by each device and the re ection against a standard-white Spectralon® offered values for calibration. The adjusted value was the raw value of the acquisition divided by the standard-white re ection.

Improvements on the gestational age estimation model
The new test has an algorithm to predict gestational age, previously described elsewhere and patented. 19,47 However, the current data set groundwork for improvements in the algorithm for predicting gestational age based on the ridge linear regression method 48 to optimize the latent collinearity and avert over tting variables. Additionally, the nonlinear machine learning method eXtreme Gradient Boosting (XGBoost) 49 created at most 50 trees with a maximum depth of three. Both models were validated using a 10-fold cross-validation approach ten-fold, repeated 30 times. Missing data had no imputation. The analytical pipeline is detailed in Supplementary Note 5.
All clinical variables used as predictors of the models were available at the test time. Therefore, they could be used in real scenarios from user input into the medical device interface. The standalone newborn skin re ectance value was the raw variable, adjusted by clinical variables. Skin re ectance had a strong positive correlation with R-gestational age; Pearson coe cient = 0.79, p < 0.001, with a mean absolute error (MAE) of 2.0 weeks (Supplementary Table in Note 2). However, skin acquisition adjusted for birth weight and antenatal use of ACTFM achieved the best performance in gestational age prediction. The set model performance is presented as a Supplementary Table in Note 2, including intermediary analysis with incubator stay, gender, and jaundice variables to verify the elimination of intervenient variables after technological improvements since the early version of the device. 19,47 Outcomes The primary outcome was the agreement between the gestational age predicted by the new test and the R-gestational age in terms of the most accurate predicted models.
A secondary endpoint was the accuracy of the new test to discriminate preterm newborns considering thresholds at 37, 32, and 28 weeks of pregnancy. Moreover, the proportion of preterm newborns correctly detected at birth within a one-week error margin. Another secondary endpoint was comparing the difference between predicted gestational age and the gestational age calculated by a second ultrasound exam after 13 weeks and 6 days of gestation and before 22 weeks via Comparator 1 (C1) and with the LMP, Comparator 2 (C2). This outcome was intended to simulate the performance of the test in scenarios without the reference and compare the agreement between established methods for gestational age calculation and the new test.
The safety of the new device is still a derived endpoint, which refers to unexpected medical events, unintended illness or injury, or unfortunate clinical signs in subjects, users, or others, whether or not they are related to the investigated product.

Statistical analysis
A sample of 787 newborns was required to determine an effect size of 10% to test the hypothesis of equivalence between gestational ages estimated by test and reference, as detailed in our research protocol. 43 Intention-to-diagnosis analysis guaranteed all newborns who were included in the statistical analysis, regardless of any result of the new test, and even inconclusive ones. Descriptive analysis of the newborn's clinical characteristics and the intervention measurements were performed.
Regarding the primary endpoint, the agreement among different methods for gestational age (R, new test, C1, C2) was calculated using the intraclass coe cient correlation and Bland & Altman plots. 50 Regarding the accuracy of predicted gestational age by the models in identifying premature newborns, the area under the receiver operating characteristic curve (AUROC) at 95% CI described the model's discrimination. The signi cance level for hypothesis tests is 5%, together with 95% con dence intervals.

Data Availability Section, Responsibility, and Analysis
The lead authors (ZSNR, RMCR, RNG, JSG and RAPLA) had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. Correspondence and requests for materials should be addressed to ZSNR. Data are available upon reasonable request and after anonymization to allow sharing of data ethically and legally, preserving the con dentiality of the persons who participated in this study.     days of gestation and before 22 weeks.
Comparator 2 is the gestational age calculated by the last menstrual period.
Model 1: Linear-Ridge skin reflectance and birth weight predictors for gestational age. del 2: XGBoost skin reflectance, birth weight, and ACTFM exposure predictors for gestational age. Figure 1 Flow diagram of participants throughout the study with results for the predictive models Notes: Model 1 Test: Linear-ridge skin re ectance and birth weight predictors for gestational age. Model 2 Test: XGBoost skin re ectance, birth weight, and antenatal corticosteroid therapy for fetal maturation exposure predictors for gestational age.   Correlation plot between the skin re ectance of the newborn and the reference gestational age at birth.

Figures
Notes: The hatched lines correspond to 95% of the con dence interval. Black diamonds represent preterm newborns by the reference gestational age. Gray circles: represent term newborns by the reference gestational age. *When Comparator 2 is available. GA: gestational age Comparator 1 is the gestational age calculated by a second ultrasound exam after 13 weeks and 6 days of gestation and before 22 weeks.
Comparator 2 is the gestational age calculated by the last menstrual period.
Model 1: Linear-Ridge skin re ectance and birth weight model predictors.

Figure 5
Receiver operating characteristic for the models to discriminate terms from preterm newborns. AUROC: the area under the receiver operating characteristic curve.
Model 1: Linear ridge skin re ectance and birth weight predictors for gestational age, n=781 Model 2: XGBoost skin re ectance, birth weight, ACFM exposure predictors for gestational age, n=777 The new device and its simulated application on a newborn-doll. Source: the authors

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download. SupplementaryMaterialUFMG.docx