Assessing placental function across gestation: a multi-institutional study of BOLD-MRI for the prediction of adverse pregnancy outcomes

19 The placenta is a remarkable organ that coordinates and regulates maternal-fetal interactions 20 during pregnancy to optimize fetal development. A host of obstetric complications are 21 associated with placental dysfunction, and existing methods for evaluating in vivo placental 22 function fail to reliably detect at-risk pregnancies prior to maternal or fetal morbidity. Although 23 routinely used as a monitoring tool, the predictive power of ultrasound for identifying 24 compromised pregnancies is poor. Recent preclinical studies performed in our laboratory, using 25 blood oxygen-level dependent magnetic resonance imaging (BOLD-MRI) in the pregnant 26 nonhuman primate (NHP), established a strong correlation between placental T2* values and 27 maternal-fetal oxygen transport. Here we extend this work to a large, longitudinal, two-site study 28 of quantitative in vivo T2* mapping in human pregnancies across 11 to 38 weeks gestation to 29 characterize the evolution of placental oxygenation in uncomplicated pregnancies and to 30 elucidate the relationship between aberrant placental T2* and adverse obstetric outcomes 31 attributable to placental dysfunction. This methodology has high discriminatory power and 32 strong potential diagnostic utility.


Introduction 35
The fundamental role of the placenta in fetal development, pregnancy morbidity, and neonatal, 36 pediatric, and even lifelong health is indisputable (1-7). Aberrant placental development has 37 been linked to virtually every adverse obstetric outcome, including abnormalities in fetal growth, 38 preeclampsia, preterm labor, and stillbirth (4, 8-17). During pregnancy, the placenta supplies 39 oxygen and critical nutrients required for fetal growth, removes waste products from the fetal 40 circulation, protects the fetus from environmental toxins and infections, produces pregnancy-41 specific hormones, and mediates communication between the fetus and the mother to 42 coordinate maternal physiologic adaptations and fetal development (18)(19)(20). The regulation of all 43 of these processes changes dynamically across gestation to ensure appropriate maternal 44 resource allocation to meet fetal growth demands. Although mechanisms regulating normal 45 placental growth and development are incompletely understood, the central role of the placenta 46 in fetal homeostasis is clear. 47 The inability to longitudinally sample placental tissue during gestation constitutes a significant 48 limitation for the study and assessment of placental development in human pregnancies. 49 Therefore, development of non-invasive tools and diagnostics to accurately characterize normal 50 development across gestation and assess placental function and health in vivo is a crucial 51 component in the identification of pregnancies at risk for adverse obstetric and neonatal 52 outcomes. 53 Obstetric imaging, predominantly with ultrasound (US), is a mainstay of clinical care for 54 identification of fetal anomalies and detection of aberrant fetal growth (21)(22)(23)(24)(25). Uterine artery 55 velocimetry has been studied as a potential predictor of preeclampsia and fetal growth 56 restriction (FGR), and has modest predictive power for severe, early onset phenotypes of both 57 (22,24,26), but it performs poorly in predicting later onset morbidity due to placental 58 dysfunction, possibly because it measures resistance to blood flow (impedance) in the umbilical 59 artery rather than focusing on perfusion of the placenta itself (27). Similarly, fetal umbilical artery 60 Doppler ultrasound is used to risk stratify pregnancies suspected to have FGR based on 61 ultrasound fetal biometry. Observation of abnormal blood flow via umbilical artery Doppler 62 assessment, particularly absent or reversed diastolic blood flow, is clearly associated with 63 adverse perinatal outcomes (28,29). However, its principal utility is in antenatal surveillance to 64 guide hospitalization and timing of delivery after the diagnosis of FGR has already been 65 established by ultrasound-based biometry, not in prediction of incipient FGR. Although it is clear 66 that profoundly abnormal umbilical artery blood flow in the setting of FGR is associated with 67 adverse perinatal outcome (28), it is not a direct measure of placental function and it can be 68 normal in some cases of severe placental insufficiency. Despite its widespread use, the utility of 69 screening ultrasound as a tool for identification of pregnancies at risk for adverse outcomes 70 remains limited. 71 Magnetic resonance imaging (MRI) has been used during pregnancy for decades, primarily to 72 assess fetal abnormalities via anatomic imaging. Recently, the NIH Human Placenta Project 73 stimulated the development and application of a number of innovative MRI techniques, intended 74 to enable in vivo assessment of placental function during pregnancy (30). Early work by 75 Sorensen and colleagues (31) observed the presence of spatial heterogeneity in T2*-weighted 76 MRI of the placenta, and found that this heterogeneity was decreased by maternal hyperoxia. It 77 is well-known that T2*-weighted images are sensitive to changes in the relative levels of 78 oxyhemoglobin and deoxyhemoglobin via the blood oxygenation level dependent (BOLD) effect, 79 which forms the basis of functional MRI (fMRI) studies of the brain (32). Consideration of the 80 specifics of the anatomy and physiology of hemochorial placentas led us to hypothesize that 81 quantitative T2* measurements could be used to assess placental perfusion and maternal-fetal 82 oxygen transport. Subsequent work performed by our group in pregnant nonhuman primates 83 6 Placental T2* 148 Figure 2 shows representative anatomic T2-weighted HASTE (left column) and quantitative T2* 149 maps (right column) acquired in two study participants, matched for gestational age at time of 150 scan. Placental regions of interest (ROIs) are superimposed on the T2* maps (blue dashed 151 lines). The upper row in the figure shows a UN pregnancy at 232 days of gestation with median 152 placental T2* (= 51 ms) close to the population median (50 th percentile), while the bottom row 153 shows corresponding images for a PA pregnancy at 235 days gestation with a median T2* (= 26 154 ms) in the 1 st percentile. Depression of the placental T2* in the latter is clearly apparent in panel 155 D. 156 Figure 3A shows the measured dependence of placental T2* across gestation in UN 157 pregnancies. This quantity decreased continuously throughout pregnancy, beginning at a 158 relatively high plateau level early in gestation, then dropping increasingly rapidly to an inflection 159 point around 30 weeks before approaching a second, lower plateau in late gestation. Model 160 fitting via nonlinear least squares regression to a logistic function is shown by the solid black 161 curve, with 95% fit confidence intervals (CI) indicated by the dashed lines and 95% fit prediction 162 intervals (PI) by the dot-dashed lines. Figure 3B plots the corresponding data and regression 163 curves for SA (green) and PA (red) pregnancies, with the best fit and 95% CI curves from UN 164 pregnancies shown in gray for reference. The model fit for the PA outcome group had 165 significantly lower modeled T2* than UN pregnancies starting at 15 weeks and continuing 166 through 33 weeks gestation, while the model fit for the SA outcome group was not significantly 167 different from that for UN pregnancies at any point in gestation. 168 Site-dependent data and regressions for UN pregnancies are shown in Figure 3C for OHSU 169 (blue) and Utah (red), with fit and 95% CI for all UN again plotted in gray. While the resulting 170 curves are quite similar in shape, the Utah T2* data for UN pregnancies are consistently lower 171 than the corresponding OHSU data, and the difference between the two is statistically 172 significant between 15 and 29 weeks gestation. The observed site-specific differences in T2* in 173 UN placentas can be accurately described with a simple model (see Methods) that 174 characterizes these differences in terms of corresponding site differences in maternal 175 hemoglobin and SpO2 levels, the known dependence of MRI signal on deoxyhemoglobin 176 concentration, and a gestation-dependent maternal placental blood volume fraction (! !"# (#)) 177 term that varies from approximately 15% early in gestation to approximately 35% by late 178 gestation. Hemoglobin and SpO2 variation between imaging sites can be explained by the 7 altitude difference, with the University of Utah at 4,840 feet above sea level while OHSU lies 180 roughly 450 feet above sea level. 181 Median voxel-level relative measurement uncertainty in placental T2* data for UN pregnancies 182 was ±7.0%, was comparable in both SA (±6.1%) and PA (±6.2%) pregnancies, and was 183 significantly higher in the Utah studies than at OHSU (±5.8% for OHSU, 10.3% for Utah, 184 p<0.001). In addition to stratifying based on pregnancy outcome and study site, the 185 dependence of gestational T2* measurements in UN pregnancies on fetal sex, maternal age, 186 and maternal body mass index (BMI) was evaluated (not plotted), with no significant differences 187 among any of these. Excluding measurements not meeting the heuristic data quality criteria 188 described in the methods did not significantly alter any reported results. 189 The average rate of change in placental T2* with gestation, computed from the centered finite 190 difference of measurements in each individual pregnancy at successive time points, is plotted 191 for UN pregnancies in Figure 3D, for PA (red), and SA (green) pregnancies in Figure 3E, and for 192 OHSU (blue) vs. Utah (red) UN in Figure 3F. Model regressions to these data using the time 193 derivative of the logistic function are displayed as in Figures 3A-3C. As with the T2* data 194 themselves, the rate of change data for UN and SA pregnancies are not significantly different at 195 any point during gestation. In contrast, the rate of change in PA pregnancies is nearly constant 196 and shows a significantly larger rate of decrease in early and mid-gestation (up to 24 weeks) 197 relative to UN. The rate of T2* decrease with gestation was found to be slightly, but significantly, 198 larger in OHSU UN pregnancies than in Utah UN from 28 weeks gestation onward. 199

Receiver operating characteristic (ROC) curves for T2* 200
Histograms of T2* z-values for uncomplicated normal (blue), primary adverse outcome (red), 201 and secondary abnormal outcome pregnancies (green) are plotted in Figure 4, where the 202 uncomplicated pregnancies were used as the reference distribution. As expected, the 203 distribution of z-scores for UN pregnancies is symmetrical and centered on zero (mean=0.0, 204 SD=1.0). Z-scores for the PA pregnancies are relatively symmetrical but broader and with a 205 significant left shift (mean=-1.0, SD=1.49, p<0.001), while the distribution for SA pregnancies is 206 shifted leftward (mean=-0.15, SD=1.34, p<0.001) and notably skewed, suggesting the possibility 207 of two subpopulations within these data. 208 The distributions of T2* percentiles derived from the z-score data are presented in bar chart 209 form in Figure 5, with twenty equally-spaced bins spanning from 0 to 100. The distribution of T2* 210 percentiles in the UN population (blue) is, as expected, essentially uniform across the entire 211 range, with roughly 5% of observations lying in each bin, while SA (green) pregnancies show 212 modest enrichment at low values. In contrast, the PA pregnancies lie primarily in the lowest (0-213 5%) bin, with nearly 35% of the adverse studies lying in that range and 44% in the lowest 10% 214 of T2* measurements. 215 Figure 6 shows receiver operating characteristic (ROC) curves for the entire population across 216 gestation (leftmost column), and for data separated into early (10-20 weeks), mid (20-30 217 weeks), and late (30+ weeks) gestation (second through fourth columns). ROCs for both sites 218 are plotted in the top row, those computed using only OHSU data in the middle row, and those 219 computed using only the Utah data in the bottom row. For both sites across all gestational time 220 points, the area under the curve (AUC) or C-statistic for placental T2* and PA pregnancy 221 outcome is 0.71, with mid-gestation showing the strongest predictive power (AUC of 0.76). C-222 statistics are consistently higher in the OHSU cohort than the Utah cohort, with the strongest C-223 statistic overall for OHSU studies in mid-gestation (AUC=0.82), and the weakest for Utah 224 studies in late-gestation (AUC=0.37). The maximum in Youden's J statistic, % !$% , is indicated by 225 the red stars on the ROC curves of Figure 6, and the corresponding optimal cutoff threshold in 226 T2* percentile relative to UN is indicated in the legends as & &"' . 227

Placental T1 228
Quantitative T1 values in UN pregnancies (acquired in OHSU participants only) showed linear 229 decrease with gestation at an average rate of -26.9 ms/week from approximately 2200 ms early 230 in gestation to roughly 1600 ms late in gestation. Neither SA nor PA pregnancies showed any 231 statistically significant differences in the evolution of T1 during pregnancy relative to UN,232 suggesting that placental T1 is not a useful metric for characterization of placental dysfunction. 233

Maternal hemoglobin and oxygen saturation 234
Maternal hemoglobin level decreased linearly throughout gestation in UN pregnancies at an 235 average rate of -0.046 mmol/week and was significantly higher in the Utah cohort than the 236 OHSU cohort (mean difference 1.04±0.90 mmol, p<0.001). Maternal hemoglobin was 237 significantly higher in the PA pregnancies compared to UN (mean difference 0.36±1.04 mmol, 238 p<0.001). There was no difference in hemoglobin between UN and SA pregnancies. 239 Maternal SpO2 in UN pregnancies was found to be essentially constant throughout gestation 240 (mean 97.0%) but was significantly lower in the Utah cohort than the OHSU cohort (mean 9 difference -2.0±3.0%, p<0.001). Neither SA nor PA pregnancies were associated with 242 statistically different maternal SpO2 values or trends relative to UN. 243

Placental volume 244
Placental volume increased linearly during gestation at an average rate of 32.2 cm^3/week 245 beginning between 11 and 12 weeks, with no significant difference between UN pregnancies at 246 the two sites. Our placental volume measurements are also highly consistent with those 247 reported by Leon et al. (46) in the overlapping gestational age range. Volume in SA pregnancies 248 was not significantly different than that of UN, while PA pregnancies showed a slightly lower rate 249 of growth (30.0 cm^3/week) leading to significantly lower volume from 20 weeks gestation 250 onward. 251

Regression modeling results 252
Model definitions, best fit parameter values, fit parameter uncertainties, and root-mean-square 253 (RMS) residual errors for regressions to all data and subsets discussed above are given in 254 Table 3. 255

Discussion 256
In this study, we characterize and model the sigmoidal evolution of T2* across gestation in 257 uncomplicated pregnancies and demonstrate that median placental T2* is markedly lower in 258 many pregnancies with adverse outcomes. We found lower values across gestation in 259 pregnancies with the primary adverse outcome and larger rate of decline in early and mid-260 gestation when compared to uncomplicated normal pregnancies. Importantly, decreased 261 median T2* continues to correlate with adverse obstetric outcomes when quantified in both mid 262 and late gestation. The placenta is a dynamic organ which evolves over the entire course of 263 gestation, and possesses the capacity to adaptively develop in concert with the growing fetus. 264 As a result, it is not a fait accompli that poor placental function early in gestation persists 265 throughout pregnancy. However, our results suggest that, while T2* measurements acquired in 266 the mid-gestational time window (20-30 weeks) are most predictive of adverse pregnancy 267 outcomes, even data from the early gestational window (10-20 weeks) allow risk stratification. 268 The strong correlation of data observed between our two independent sites demonstrates that 269 this method is robust and has the potential to be transferable across different institutions. 270 Nevertheless, some relevant site-specific differences were observed that merit further 271 clarification. First, the birthweight in PA pregnancies among the Utah group was not 272 significantly different from that of the UN pregnancies in that group, while the PA pregnancies in 273 the OHSU cohort had a significantly lower birthweight when compared to UN. This may simply 274 be accounted for by the higher proportion of SGA neonates in the OHSU cohort due to chance, 275 as population rates of adverse outcomes in these two groups are expected to be similar. 276 Second, the maternal Hgb was higher in Utah when compared to OHSU, which is expected 277 given the increased altitude in Salt Lake City, Utah when compared to Portland, Oregon. Third, 278 maternal oxygen saturation in the Utah patients was significantly lower than in the OHSU 279 population, also consistent with the physiologic impacts of altitude. In the UN population, it is 280 possible to entirely explain the observed differences between sites with a simple model 281 incorporating the site-specific hemoglobin and oxygen saturation differences along with a 282 maternal placental blood volume term that varies across gestation, as described in the Methods. 283 The predictive power of T2* measurements for discriminating uncomplicated pregnancies from 284 primary adverse outcome pregnancies was much higher in the OHSU cohort than for Utah 285 (AUC 0.80 vs 0.56). We suspect that this is due to site-specific differences in the prevalence of 286 SGA and preeclampsia with severe features, both of which are relatively under-represented in 287 the Utah group. Notably, in the Utah cohort, birthweights of neonates in the primary adverse 288 outcome group were not statistically different than in uncomplicated pregnancies. Given that 289 SGA and hypertensive diseases of pregnancy have multiple pathophysiologies with varying 290 degrees of placental dysfunction, it is reasonable to propose that T2* quantification is primarily 291 predictive of pathways linked to abnormalities attributable to perturbations of maternal placental 292 blood flow and/or fetal oxygen uptake. It is possible that there is a secondary contribution due to 293 the somewhat higher measurement error in the Utah data set as compared to OHSU, although 294 the absolute measurement uncertainties are small for both study sites. Unfortunately, the 295 modest number of PA pregnancies in our data set limits statistical power and precludes 296 separation of the PA group into sub-categories. 297 The imaging methodology in this study is highly amenable to clinical translation. Placental MRI 298 was performed using imaging protocols and pulse sequences that are available on virtually all 299 modern MRI scanners, and analysis of these data requires only minimal post-processing to 300 convert signal measurements to T2* values. Groundwork performed in our NHP models was 301 key to both validation and translation of this methodology to human subjects by validating T2* 302 mapping as a functional measure of maternal placental perfusion with confirmation by 34). While DCE-MRI is the gold-standard method for quantifying tissue perfusion via MRI, 304 and despite the fact that we have demonstrated minimal placental permeability to passage of 305 gadolinium-based contrast agents (GBCA) following in utero maternal administration (47, 48), a 306 GBCA-free alternative alleviates potential reservations to use of MRI as a clinical diagnostic 307 imaging tool for assessing placental health. In addition, because placental T2* is sensitive to the 308 balance between oxygenated maternal blood delivery and fetal oxygen demand, it is particularly 309 well-suited to identify problems stemming from inadequate placental oxygenation. 310

Study strengths and limitations 311
Our study has a number of strengths. It is the largest prospective study of placental MRI, and 312 the most extensive study of T2*, in particular. In addition, the longitudinal design enabled us to 313 characterize the nonlinear evolution of T2* across pregnancy and provide reference values for 314 both T2* itself and rate of change in T2* as gestation progresses. While studies of changes in 315 T2*-weighted BOLD-EPI measurements in response to hyperoxygenation have a number of 316 advantages, particularly in data acquisition efficiency and sensitivity to motion, they are 317 generally semi-quantitative, introduce methodological complexity, and potentially perturb 318 maternal and fetal hemodynamics and alter the physiologic mechanisms that determine normal 319 oxygen transport across a gradient (49-51). In contrast, quantitative measurements of T2* are 320 reflective of the balance between maternal delivery of oxygen and fetal demand, are 321 reproducible across sites, and do not need ancillary experimental perturbations. The primary 322 adverse composite outcome metric we developed was defined prior to, and independent of, MRI 323 data analysis. Designation of pregnancy outcome was blinded to MRI data analysis and was 324 conducted by four Maternal-Fetal Medicine physicians independently, increasing the rigor of our 325 outcome designation. Similarly, to further reduce the potential for bias, MRI data processing 326 was blinded to pregnancy outcome and was conducted independently prior to statistical analysis 327 for association with adverse pregnancy outcomes. By utilizing common, commercially available 328 MRI acquisition protocols, the work described here should be easy to reproduce at other 329 institutions, facilitating its potential use both in future clinical studies and in clinical practice. 330 There are also a number of limitations to this study. Although it is the largest longitudinal MRI 331 study in pregnancy performed to date, the number of adverse outcomes was small. This 332 necessitated the utilization of a composite outcome, as is typical for obstetric studies. Our study 333 population is relatively ethnically and racially homogeneous, so the conclusions we draw here 334 may not be applicable to other populations. MRI was performed using 3 Tesla scanning 335 hardware to increase sensitivity to changes in T2*, but these systems are not currently the 336 standard in obstetric imaging and are not as widely available as 1.5 Tesla systems. While we 337 used consistent criteria encompassing many common complications, there is no gold standard 338 definition of placental dysfunction, and our outcomes are heterogenous by nature. In particular, 339 we have previously identified circumstances where pathology related to villous inflammation or 340 malformation can cause elevated T2* (35) in the setting of adequate supply of maternal arterial 341 blood to the placenta in conjunction with impaired trans-villous oxygen permeability, which could 342 constitute a confounding factor in some pregnancies. As a result, further refinement may be 343 required to detect abnormally high, as well as abnormally low T2*, to accurately capture 344 different types of placental pathology. 345

Conclusion 346
We present the results of a prospective longitudinal human study that demonstrate the potential 347 of quantitative T2* mapping during pregnancy to identify increased risk for adverse obstetric 348 outcome due to placental dysfunction, particularly in the setting of fetal growth restriction. 349 Quantitative measures of placental T2* identified pregnancies at increased risk for adverse 350 outcomes across all gestational ages in this study despite site-specific differences in maternal 351 and neonatal demographics at the two institutions. Low median placental T2* was strongly 352 correlated with low fetal birthweight, suggesting that the diagnostic utility of placental MRI may 353 be enhanced by focusing on a specific adverse obstetric outcome, such as fetal growth 354 restriction, rather than a composite adverse outcome. Improved diagnostics to identify 355 pregnancies at risk of adverse outcomes may facilitate discovery of novel biomarkers, improved as some participants recruited later in gestation were not be able to complete additional MRI 378 studies depending on the gestational age at enrollment. The decreased frequency of repeat MRI 379 per study subject however did facilitate recruitment and enrollment of a larger study cohort than 380 originally planned. 381 Participants were recruited from the OHSU and UUHSC clinics where written informed consent 382 was obtained with IRB approval. Pregnant women were recruited based on inclusion criteria for 383 two subject groups: a low-risk cohort not at increased risk for adverse obstetric outcome and a 384 high-risk group at increased risk for adverse outcomes based on prior clinical history. A third 385 group of pregnant tobacco smokers was originally planned as a separate cohort but recruitment 386 was abandoned due to lack of success in identification and enrollment. 387

Inclusion criteria 388
Inclusion criteria for both groups included pregnancy (defined by positive pregnancy test and 389 certain menstrual history, or early ultrasound) identified prior to 16 weeks gestation, maternal 390 age over 18 years of age, and ability to give informed consent. The inclusion criteria for the low-391 risk cohort were all of the following: 1) no history of a second or third trimester pregnancy loss, 392 2) no history of fetal growth restriction, and 3) nonsmoker. The inclusion criteria for the high-risk 393 group were one or more of the following: 1) history of pregnancy complicated by placental 394 insufficiency in a previous singleton pregnancy defined by preeclampsia with severe features 395 requiring preterm delivery, or preterm delivery due to placental insufficiency (fetal growth 396 restriction, oligohydramnios, abnormal umbilical artery Doppler), or fetal growth restriction with 397 neonatal weight < 10 th percentile delivered at term, or stillbirth attributed to placental cause, 398 regardless of gestational age, 2) pregnancy at risk for placental insufficiency due to clinical 399 comorbidities (i.e. chronic hypertension), or 3) history of spontaneous preterm birth < 34 weeks. 400

Exclusion criteria 401
Exclusion criteria were maternal intellectual disability or incarceration, pregnancy with major 402 fetal anomalies known to be associated with abnormal fetal growth, active alcohol use during 403 pregnancy, medical conditions requiring ongoing treatment during pregnancy including cancer, 404 acute liver disease, chronic pulmonary disease requiring regular use of medication, history of 405 claustrophobia, metal implants, or other contraindication for MRI, and increased risk of 406 aneuploidy based on ultrasound findings and/or genetic testing. 407

Participant enrollment 408
Potential participants were identified through multiple modalities. The research team utilized 409 social media, which entailed Facebook advertisements and promotions via institutional 410 websites. The research teams also attended multiple pregnancy groups, such as prenatal group 411 intake meetings and events for pregnant women, as well as a Portland-based website for new 412 and expecting parents with resources and events. Potential research subjects were screened in 413 the OHSU electronic medical record system, and Utah appointment logs through a waiver of 414 authorization. Once participants were found via electronic medical records, they were 415 approached at their next prenatal appointment or sent a MyChart message with pertinent 416 research study information. When a potential subject reached out to the team via phone or 417 email with interest, a phone screening was conducted. The phone screening reviewed basic 418 eligibility inclusion and exclusion criteria, contact information, and additional preferences. If the 419 subject was found to be eligible, they would be scheduled for a visit in accordance with the 420 study protocol where they will start the visit with a detailed explanation of the study and followed 421 with the signature of the informed consent. If not eligible for the study or no longer interested, 422 they would be thanked for their time and interest. 423

Pregnancy outcome designation 424
Pregnancies were categorized postnatally into three groups: a) uncomplicated normal 425 pregnancies (UN), b) primary adverse outcome pregnancies (PA), and c) secondary abnormal 426 pregnancies (SA). Uncomplicated pregnancies were defined as those with term (37 weeks or 427 beyond) delivery with birthweight between the 5 th and 95 th percentile, without gestational 428 hypertensive disease, and not meeting any criteria for the primary adverse outcome or 429 secondary abnormal outcome. 430 The primary adverse outcome group was defined using a composite including hypertensive 431 disorders of pregnancy, with birthweight below the 5 th percentile by Oken (43), and stillbirth or 432 fetal death. Hypertensive disorders of pregnancy included gestational hypertension, 433 preeclampsia without severe features, preeclampsia with severe features, Hemolysis, elevated 434 liver enzymes, and low platelet count (HELLP) syndrome, and eclampsia. These were defined 435 by ACOG criteria (52). Gestational hypertension was defined as systolic blood pressure of 140 436 mm Hg or more or a diastolic blood pressure of 90 mm Hg or more, or both, on two occasionas 437 at least 4 hours apart at or beyond 20 weeks of gestation in someone without chronic 438 hypertension. Preeclampsia was defined as elevated blood pressure (as in gestational 439 hypertension) plus proteinuria (300mg or more in a 24 hour urine collection or urine 440 protein/creatinine ratio of 0.3 or more) or presence of severe features without proteinuria. 441 Severe features included systolic blood pressure of 160 mm Hg or more or diastolic blood 442 pressure of 110 mm Hg or more on two occasions at least 4 hours apart, thrombocytopenia with 443 platelet count < 100 x 10 9 /L, liver enzymes more than twice the upper limit of normal, severe 444 persistent right upper quadrant pain or epigastric pain, serum creatinine > 1.1 mg/dL, pulmonary 445 edema, new headache unresponsive to medication, and/or visual disturbances. HELLP 446 syndrome is defined by thromboycytopenia with platelet count < 100 x 10 9 /L, liver enzymes 447 more than twice the upper limit of normal, and evidence of hemolysis (LDH > 600 IU/L). 448 Eclampsia was defined as seizure in absence of alternative etiology and with concomitant 449 criteria for preeclampsia. 450 The secondary abnormal outcome group included pregnancies complicated by maternal chronic 451 hypertension without superimposed preeclampsia, fetal genetic and/or anatomic anomalies; 452 spontaneous preterm birth due to preterm labor, cervical insufficiency, and/or preterm 453 premature rupture of the membranes (PPROM); placental abruption, chorioamnionitis, and/or 454 birthweight greater than the 95 th percentile by Oken (43). Our goal in creating this group was to 455 capture pregnancy morbidity that is typically due to non-placental causes. Though morbidity can 456 be significant in this group, our hypothesis was that our MRI protocol would not show strong 457 association with these outcomes. 458 The primary adverse clinical composite outcome was developed pragmatically. Gestational 459 hypertensive disease, preeclampsia, low birth weight, and fetal death/stillbirth are all linked to 460 placental dysfunction (3, 50, 51). We acknowledge that this is not a pure phenotype, and there 461 are multiple pathways to each of these outcomes, let alone the composite. However, this 462 composite is clinically meaningful when attempting to capture morbidity and mortality due to 463 placental insufficiency and there is significant overlap between the individual outcomes. Thus, in 464 order to assess utility of MRI as a tool for either prediction or targeting of therapeutics, 465 assessing these outcomes in toto is appropriate. 466 Adjudication of the primary adverse outcomes and secondary abnormal outcomes was 467 performed independently by both of the Maternal-Fetal Medicine physicians at each site (OHSU: 468 KJG, AEF; Utah: JMP, NRB). Any discordance between assessment of outcome was then 469 discussed and reconciled prior to final determination. The authors determining the outcomes 470 were blind to the MRI data and analysis. Similarly, MRI data acquisition and quality control was 471 performed blind to pregnancy outcome group. 472

Magnetic Resonance Imaging 473
Maternal blood draws were performed prior to each scan and hemoglobin level measured using 474 iStat (Abbott, Princeton, NJ) and/or fingerstick. Pulse oximetry (Zacurate 500BL) was used to 475 determine maternal blood oxygen saturation level before each MRI study. Three dimensional placental regions-of-interest (ROIs) were hand drawn on T2* maps, using 497 co-registered T2-HASTE images as an anatomic reference when the placental-myometrial 498 and/or placental-fetal boundaries were indistinct. Slices with excessive image quality 499 degradation arising from maternal and/or fetal motion were excluded from analysis. Readers 500 were blinded to patient status and pregnancy outcome. Binary masks were derived from the 501 placental ROIs, with T2* values of 250 ms or more being excluded from further analysis; such 502 large values are associated with signal contamination by amniotic fluid. Placental volume was 503 computed by summing the number of voxels in each slice of this binary mask multiplied by the 504 per-voxel volume. Where slices were missing due to motion, volumes were estimated from 505 adjacent slices using linear interpolation. In the OHSU cohort, median placental T1 was 506 determined by spatially resampling measured T1 maps onto the T2* image volumes and 507 applying placental ROIs. 508 Histogram analysis was used to compute median placental T2* for each study. Median relative 509 fit uncertainty in T2* (the ratio of the model estimated sigma-T2* to estimated T2*) was 510 computed as a measure of measurement quality. A heuristic quality statistic was determined by 511 computing the fraction of placental voxels for which the relative fit uncertainty was <= 0.25. 512 Sensitivity of the data analysis pipeline to data quality was evaluated by re-running the entire 513 statistical post-processing pipeline for both the entire set of studies and for a reduced set where 514 studies with either median relative uncertainty in the highest 10% or heuristic quality in the 515 lowest 10% were excluded. Potential bias stemming from variation in the number of scans per 516 patient ranging from one to three was considered by re-running the post-processing pipeline 517 with scan weight distributed evenly per patient, rather than per scan. 518 Trends in maternal SpO2, maternal hemoglobin concentration, placental volume, and placental 519 T1 across gestational age were found to be well fit with a linear model. Gestational trends in T2* 520 were modeled sequentially with linear and quadratic polynomials as well as with a logistic 521 function (' ( * (#) = ) * /(1 + exp0) ( (# − ) + )2) + ) , ), which was chosen based on the sigmoidal 522 behavior observed in median T2* curves in normal pregnancies. Regression modeling used 523 polyfit for polynomial fitting and nlinfit for nonlinear model fitting to T2* data, with 524 regressions weighted by measurement uncertainties. The Bayes Information Criterion (BIC) was 525 used to account for different numbers of free parameters for model selection. Based on the BIC 526 results, the sigmoid function was determined to have the best fit to measured T2* data. Intra-527 individual time derivatives in T2* across gestational age were approximated as pairwise 528 centered finite differences from measurements acquired at sequential gestational time points: 529 Fit covariance matrices were used to compute model parameter uncertainties as well as 95% 532 confidence and prediction intervals for all regressions. Model regressions were considered to be 533 significantly different where their 95% confidence intervals were non-overlapping. In addition to 534 the three primary study groups (UN pregnancies,SA pregnancies,and PA pregnancies),535 differences between UN at the two study sites were compared, along with differences based on 536 fetal sex, maternal age, and maternal BMI. 537 A model describing the observed inter-site T2* differences across gestation between OHSU and 538 Utah sub-populations was developed starting from the assumption that these differences arise 539 from corresponding site-dependent differences in maternal hemoglobin and SpO2 levels, along 540 with a maternal placental blood volume fraction that evolves through pregnancy. The resultant 541 shift is most easily described in terms of differences in the transverse relaxation rate, defined as 542 6 ( * = 1/' ( * , between site A (OHSU) and site B (Utah) using the known transverse relaxivity of 543 deoxyhemoglobin (7 ( * =20.2/mmol/s) and blood deoxyhemoglobin concentrations computed from 544 measured maternal SpO2, along with the maternal placental blood volume fraction (! !"# ): 545 Combining this expression with the modeled dependence of maternal SpO2 and hemoglobin 547 levels for the two sites allows us to solve for the gestation-dependent value of ! !"# (#) that 548 corresponds to observed differences in placental T2*. This parameter is a volume fraction that 549 should correspond approximately to in vivo intervillous volume (55), and is physically-550 constrained to lie in the closed interval [0,1]. 551 Z-scores were computed for all placental T2* measurements using the model regression and 552 prediction intervals for UN pregnancies, and T2* percentiles were calculated from the 553 cumulative distribution function (CDF) of the corresponding normal distribution. Receiver 554 operating characteristic (ROC) curves were then generated from the T2* percentiles for all 555 individual studies in UN vs. PA pregnancies (SA pregnancies were omitted from this analysis). 556 Studies were also grouped into three gestational time windows: early gestation (10-20 weeks), 557 mid-gestation (20-30 weeks), and late gestation (30-40 weeks) in order to assess the 558 performance of this metric across gestation, and ROCs were separately computed for OHSU 559 and Utah data to compare the site-specific performance. 560

Figure 2
Comparison of anatomic imaging (T2-weighted HASTE, left column) and placental T2* mapping (right column) in a uncomplicated normal pregnancy at 232 days gestation (top row, panels A & B) with those from a primary adverse pregnancy at 235 days gestation presenting with severe preeclampsia (bottom row, panels C & D). The placenta is indicated by the dashed blue outlines overlaid on the T2* maps.

Figure 3
Gestational dependence of placental T2* values. Median T2* values for each completed study, computed over the entire placenta, are plotted as a function of gestational age at time of imaging in the three panels in the left column (panels A, B, C), while corresponding rates of change in placental T2* between repeated imaging time points within the same individual are plotted as a function of gestational age in the right column (panels D, E, F). The upper row plots these quantities for normal pregnancies, the middle row for abnormal (green) and adverse (red) pregnancies, and the bottom row for normal pregnancies strati ed by site (OHSU in blue, Utah in red). In all graphs, model regression curves (using the functions de ned in Table 2) are indicated by the thick solid lines, the 95% con dence intervals by the dashed lines, and the 95% prediction intervals by the dot-dashed lines. The best t and 95% CI curves from the plots in the upper row are shown in gray in the middle and bottom rows for reference.

Figure 4
Histograms of T2* z-scores in normal, abnormal, and adverse pregnancies. Z-score histograms shown are computed using prediction intervals for sigmoid model regression to T2* measurements in UN pregnancies, applied to individual studies in UN (blue), PA (red), and SA (green) pregnancies.

Figure 5
Bar chart of distribution of measured T2* percentiles for uncomplicated normal, primary adverse, and secondary abnormal pregnancies.

Figure 6
Receiver operator characteristic (ROC) curves for T2* measurements in pregnancies with our primary adverse outcome relative to uncomplicated normal pregnancies. The points where Youden's J is maximized are indicated by the stars. Area under the curve (AUC), J max, and the corresponding optimal cutoff threshold in T2* percentile relative to UN (C opt) are given in the gure legend for each panel.