Machine learning identification of maternal inflammatory response and systemic inflammatory response from placental membrane whole slide images

Abstract

Introduction

The placenta forms a critical barrier to infection through pregnancy, labor and, delivery. Acute placental inflammation in the membranes, diagnosed as Maternal Inflammatory Response (MIR), has short-term, and long-term consequences for offspring health. Digital pathology and machine learning can play an important role in understanding placental inflammation. The goal of this work is to develop machine learning models to predict pathologist-classified MIR stages from based on whole slide images (WSI) and establish early benchmarks.

Methods

We used digitized histopathology whole slide images (WSI), the UNI foundation model and the Multiple Instance Learning (MIL) framework to predict MIR from WSI images. To assess the relationship between local and systemic inflammation, we also attempt prediction of white blood cells count (WBC) and maximum fever temperature (Tmax). To analyze model performance, we built a classifier to identify inflammation in individual patches extracted from whole slide images.

Results

We were able to classify MIR with a balanced accuracy of up to 88.4% with a Cohen’s Kappa (κ) of up to 0.758. We found that deviations from the pathologist’s diagnoses were driven unusually low or high frequencies of patches with inflammation. For WBC and Tmax prediction, we found mild correlation between actual values and those predicted from histopathology WSIs.

Discussion

MIR staging is reasonably well predicted using MIL on WSIs. We show a weak but statistically significant correlation between histologic findings and maternal systemic inflammation. These findings could be used to improve reliability of classification and prediction of neonatal outcomes.

Introduction

Maternal inflammatory response (MIR)

Maternal inflammatory response (MIR) occurs when maternal leukocytes, primarily neutrophils, enter placental tissues. Classically, this infiltration is thought to respond to ascending bacterial infection,^1,2 but may also represent sterile inflammation in response to bacterial products or forces of labor.^3,4 MIR is mechanistically and statistically associated with clinical chorioamnionitis, and could be considered the localized tissue component of a systemic maternal inflammatory response, however they are not synonymous.^5–7 MIR is associated with several adverse health outcomes. Bacteria causing MIR may enter the maternal or fetal circulation, causing puerperal fever or early onset neonatal sepsis, respectively. When MIR co-occurs with FIR, there is an increased risk of neonatal death.^8,9 MIR2 is a risk factor for recurrent wheeze^10,11, asthma^10,12,13, and chronic lung disease.¹³ MIR has also been associated with lower mental development index¹⁴, and increased risk of autism spectrum disorder.¹⁵ In a small case-control study, severe MIR has also been linked to cerebral palsy albeit indirectly.¹⁶

Maternal Inflammatory Response: Definitions and Pathogenesis

Acute Placental Inflammation (API) is categorized into Maternal Inflammatory Response (MIR) or Fetal Inflammatory Response (FIR) depending on the origin of the inflammatory cells. MIR is characterized by extravasation of maternal neutrophils with movement towards the chorionic layer, followed by moving across the amnion and into amniotic space. MIR is divided into 3 stages. These stages are defined according to the Amsterdam criteria.¹⁷ Stage 1 corresponds to subchorionitis, where maternal neutrophils are limited to the cellular chorion in the membranes and subchorionic fibrin in the chorionic plate. Stage 2, chorioamnionitis is characterized by neutrophilic migration into the fibrous chorion and amniotic connective tissue. Stage 3, chorioamnionitis with amnion necrosis, is characterized by presence of neutrophil karyorrhectic debris, abscess formation, and thickened basement membranes. MIR is classically seen in the presence of ascending infection by Group B Streptococcus, E. coli and other pathogenic bacteria.^18,19 However, organisms are not always identified and some authorities argue that MIR may be the result of sterile inflammation.^{3,5,7,20–22} MIR is often seen in the setting of maternal systemic inflammation – clinical chorioamnionitis, tripe-I or intraamniotic infection. However, the association is relatively loose, MIR may be seen without evidence of systemic inflammation and vice versa.^1,6,21,22

Challenges in analyzing MIR

The stages of MIR are broadly defined with significant variability in how different pathologists classify these patterns and the cut-off between different stages^23,24. In a study by Redline et al.²⁴, showed a moderate kappa agreement of 0.58 for the diagnosis of no MIR (MIR0) vs. any MIR (MIR1–3). An older study using a previous definition of histologic chorioamnionitis showed a kappa of 0.72.²⁵ A study using secondary review of generalist pathologist diagnoses by an expert placental pathologist showed a kappa for any MIR of 0.6.²⁶ This variability creates challenges for analysis.

Machine Learning and Multiple Instance Learning

Machine Learning along with digital pathology allows investigations to be conducted with larger datasets.^27–29 Consequently, there have been several studies conducted in recent years utilizing machine learning to build tools for pathological diagnoses, and analyze digital pathology datasets. These include both local-level tasks like nuclei segmentation, and global-level tasks prostate cancer detection.³⁰

Whole slide images may exceed 100,000 × 60,000 pixels– too large for efficient computation. Multiple instance learning (MIL) based methods were proposed, that take all the patches into account, and learn to ascribe importance to patches. The importance scores can be calculated using probability of belonging to the positive class,³¹ as a proxy for selecting highly important patches. Another approach was proposed by Lu et al.³², where attention mechanism was used to force the model to ascribe importance to patches during classification. These attention scores were calculated based on features extracted using some pretrained model e.g., ResNet50. Recently, several groups have released pathology foundation models. These models use transformer architecture and have been trained on patches from tens of thousands to millions of slides.^33–35 These models show superior performance on benchmarks. However, the training sets lack placental slides and benchmarks are nearly all on detection, classification, and mutation identification of neoplasia.

Recently, several studies have been conducted employing machine learning and MIL for predicting and analyzing placenta.³⁶ Chen et al.³⁷ used CNNs to design a system for morphological characterization, and clinically meaningful feature analysis of placentas from photos. Zhang et al.³⁸ propose a Cycle-GAN along with attention module and saliency constraint to enable cross-domain image segmentation by translating target domain placenta pictures (from 1 hospital) into source domain (a different hospital), and adapting a pretrained segmentation model to segment them. Clymer et al.³⁹ designed multiresolution CNNs to classify decidual vasculopathy in placental membranes. Mobadersany et al.⁴⁰ used MIL framework to improve prediction of gestational age from placental WSI. Andreasen et al.⁴¹ proposed deep learning method for placenta segmentation from obstetric ultrasound. Goldstein et al.⁴² used the MIL framework for classification of placental villous infarction, perivillous fibrin deposition, and intervillous thrombus. Patnaik et al.⁴³ used a pretrained ResNet-18 for feature extraction from placental histopathology images, and classify maternal vascular malperfusion. Irmakci et al.⁴⁴ investigated the challenges posed to Machine Learning models by presence of tissue contaminants e.g., floaters in WSIs. Studies have also been conducted to classify cell and regions in the placental disc.^45,46

Recently, Chou et al.⁴⁷ used machine learning and other quantitative techniques to characterize maternal inflammatory response (MIR) in placental membranes.

The goal of this work is to develop machine learning models to predict pathologist-classified MIR stages from based on whole slide images (WSI) and benchmark performance against human interrater agreement.

Methods

Dataset: Patients and Placentas

The study was approved by our Institutional Review Board as STU00214052 and operated under waiver of consent. Inclusion criteria were singleton placentas examined at our institution between 2010 and 2024 with a single slide containing only membrane roll. We selected cases with and without MIR for retrieval and scanning. Slides were digitized on a Leica GT450 scanner with a 40× objective magnification (0.263 μm per pixel). Patient demographics, laboratory, and placental pathology information were extracted from our electronic data warehouse. Slides were linked to pathology reports. MIR stage involving the membranes was identified from reports using regular expressions.⁴⁸ Our site stages but does not grade MIR. Highest fever (Tmax) was defined as the highest maternal temperature recorded in the 72 hours before delivery. White blood cell count (WBC) was defined as the highest maternal white blood cell count measured in the 72 hours before delivery.

Data processing for MIR prediction:

One slide of membrane roll was digitized per patient. Patches of size 224 × 224 with no overlap were extracted at 20x from WSI of placental membranes from each patient. We used UNI³⁴ as a feature extractor. UNI³⁴ is a pathology foundation model, trained on diverse datasets of histology images, though not including placenta. These features were stored in ’tfrecords’ files for the aggregation and classification step in MIL pipeline. We analyzed the multiclass classification performance of the model, where all the MIR stages were predicted. Furthermore, we also analyzed the model performance by simplifying the MIR stage classification problem by combining data from MIR0 and MIR1 into one class, and MIR2 and MIR3 into another, and training a separate model for this task.

Dataset was split into training, validation and testing splits, where 20% of the data (678 cases) were held-out as the test set, and the training (2436 cases) and validation set (271 cases) were used for training and hyperparameter optimization respectively.

Data processing for white blood cell count and fever prediction:

For white blood cell count and Tmax prediction, we used UNI for extracting features which were then aggregated using learned attention scores, and used by the regression network for prediction. The data split between training, validation, and test was created in a similar manner as for MIR prediction. However, there are fewer data point available with WBC (1320 training, 142 validation, and 370 test cases) and Tmax (2256 training, 251 validation, and 632 test cases) information.

Machine Learning Models

We used an attention-based multiple instance learning architecture similar to that proposed by Lu et al.³². The architecture splits information processing into two stages- First stage involves patch-level feature extraction using a pretrained network. The pretrained network filters information from patches, that is available during stage 2 i.e., aggregation for the final prediction. Thus, the representations learnt by feature extractor model can play a crucial role in the final predictions. Stage 2 involves aggregation of patch-level feature to make the final prediction. The architecture uses attention to weigh different patch features. These attention weights are themselves learned during training.

Foundation model for feature extraction:

As a first step for the MIL pipeline, we extracted patch-level features from each slide using UNI as our feature extractor. UNI is a widely regarded pathology foundation model and is expected to extract relevant features for downstream classification task.

Loss function:

For MIR stage prediction we used the sparse categorical cross-entropy loss (keras). To mitigate class imbalance, we applied inverse-frequency class weighting during training.

The weight assigned to class c was computed as

$$w_c=1+\frac{N}{N_c}$$

where N is the total number of training cases and $N_c$ is the number of cases in class c

For white blood cell count prediction and fever prediction, we used mean squared error (mse), and weighted mean squared error (wmse). For weighted mean squared error calculation, we first split the ground truth values into 10 bins, then assigned weights to these bins which are inversely proportional to the number of cases in the corresponding bin. The mean squared error corresponding to each case is then weighted according to the bin its ground truth belongs to.

Optimization and regularization:

All models were trained using the AdamW optimizer with an initial learning rate of 5× 10⁻⁵. To mitigate overfitting, instance-level dropout was applied during training, randomly dropping 25% of training patches per iteration. In addition, weight dropout with a rate of 0.2 was applied to model parameters as a regularization strategy.

Learning rate scheduling and early stopping:

The learning rate was adaptively adjusted using a ReduceLROnPlateau scheduler based on the validation softmax loss. When the validation loss failed to improve for three consecutive epochs, the learning rate was reduced by a factor of 0.5, with a minimum learning rate of 10⁻⁸. Early stopping was employed with a patience of 20 epochs, and the model parameters corresponding to the lowest validation loss were restored.

Patch annotated:

We annotated 15,389 224×224 patches extracted at 20x from the MIR dataset. Patches were labeled in the single most relevant category as amnion, bacteria, blood/fibrin, decidua, decidual necrosis, trophoblast layer, fibrous tissue, glass/blank, consistent with MIR1 (cwMIR1), cwMIR2, cwMIR3, necrosis/debris, or incidental villi. A multilayer perceptron classifier with 100 neurons was trained using the UNI features. Inference was performed across the MIR test set.

Evaluation Metrics

For MIR prediction, we used Balanced Accuracy, Mathew’s Correlation Coefficient (mcc), and Cohen’s Kappa (κ) to evaluation the classification task. To evaluate white blood cell count prediction and fever prediction, we used mean squared error (MSE), mean absolute error (MAE), and R²-score.

Results

Table 1 summarizes the information about patients. Patients with MIR0,1 and MIR2,3 were of similar maternal age at delivery, and had a similar gestational age, although with a higher variation. Patients with MIR2,3 were more likely to be nulliparous, and less likely to have diabetes and hypertension in pregnancy, than patients with MIR0,1.

Table 1.

The median (IQR) values for maternal age, gestational age, Tmax and WBC are shown below, along with the count (percentage) for parous, diabetes, and hypertension. p-values are reported from Kruskal–Wallis test for quantitative values and Chi-squared test for categorical values.

	no MIR	MIR1	MIR2	MIR3	p
n	2018	206	335	156
Maternal age (years)	33.0 (30.0 – 37.0)	32.0 (28.0 – 35.0)	32.0 (28.0 – 36.0)	32.0 (27.8 – 35.2)	2.24E-07
Gestational age (weeks)	37.7 (34.7 – 39.3)	38.9 (33.9 – 39.9)	39.0 (32.3 – 40.3)	28.3 (22.9 – 38.5)	3.68E-21
Tmax (degrees C, n = 2008, 203, 332, 156)	37.2 (36.9 – 37.4)	37.4 (37.2 – 38.1)	37.7 (37.2 – 38.4)	37.7 (37.3 – 38.3)	1.85E-69
WBC (k/μl, n = 1166, 112, 195, 115)	9.9 (8.8 – 14.1)	10.1 (9.0 – 15.7)	13.4 (9.4 – 18.4)	16.3 (10.9 – 21.4)	2.13E-19
Parous	999 (0.5)	59 (0.29)	99 (0.3)	66 (0.42)	1.48E-08
Diabetes in pregnancy	253 (0.13)	19 (0.09)	25 (0.07)	19 (0.12)	0.058
Hypertension in pregnancy	598 (0.3)	28 (0.14)	41 (0.12)	24 (0.15)	9.04E-12

MIR stage prediction: Balanced accuracy, MCC, Cohen’s kappa

The multiclass classification model achieved an overall accuracy of 81% and a balanced accuracy of 63%, Cohen’s kappa (κ) of 0.557, and Matthew’s Correlation Coefficient (mcc) of 0.559. Table 2 shows the actual stage and the predicted stage for the multiclass classification model. We also analyzed the model performance by the binary classification problem (MIR01 vs MIR23). The model achieved an overall accuracy of 93% and balanced accuracy of 88.4%, Cohen’s kappa (κ) of 0.758, and Matthew’s Correlation Coefficient (mcc) of 0.758. Table 3 shows the actual stage and predicted stage for binary classification model.

Table 2.

MIR actual stage vs predicted stage for multiclass classification model

MIR	Pred-0	Pred-1	Pred-2	Pred-3
Label-0	1843	140	32	3
Label-1	65	79	57	5
Label-2	50	84	136	65
Label-3	1	8	19	128

Table 3.

MIR actual stage vs predicted stage for binary classification model

MIR	Pred-0/1	Pred-2/3
Label-0/1	2118	106
Label-2/3	91	400

Investigating model failure

To understand how the model functions or malfunctions, we investigated a case where the binary model gives false-positive i.e. where a MIR01 is predicted as MIR23. The attention heatmap, and top-5 attention patches are shown in Fig. 2. We see that the top 2 highest attention patches show decidual necrosis, and 1 patch shows mild chorioamnionitis.

Figure 2.

Attention heatmap in a False positive case for the binary classification model. The actual stage is MIR01 but is predicted as MIR23. The top 5 attention patches show 2 patches with decidual necrosis and 1 with mild chorioamnionitis.

To more systematically understand what drives model decisions, we trained a patch classification model to assign labels of consistent with MIR (cwMIR1, cwMIR2, cwMIR3) to patches in our test dataset. Then, we analyzed the percentage of cwMIR1, cwMIR2, cwMIR3 patches in each case stratified based on whether it is predicted to be of a lower stage, correct stage, or higher stage than the pathologist diagnosed (Figure 3).

Figure 3.

The percentage of cwMIR1, cwMIR2, and cwMIR3 patches in cases of patients with a) no MIR, b) MIR 1, c) MIR 2, d) MIR 3, stratified by whether the model was accurate, predicted a stage lower than diagnosed, or higher than diagnosed.

In patients without MIR (Fig. 3a) where the model predicts MIR is present, there is a higher frequency of cwMIR patches, particularly cwMIR1. Case with a pathologist diagnosis of MIR1 and model prediction of MIR0 show lower frequencies of cwMIR1 patches. Conversely, MIR1 cases predicted as MIR2 or MIR3 tend to have higher proportions of cwMIR1. In patients with MIR2, misdiagnosis as MIR1 or 0 is driven by low cwMIR2 counts while misdiagnosis as MIR3 is associated with frequent cwMIR3 patches. In MIR3 patients, misdiagnoses as MIR2, 1, or 0 seems driven by a lower frequency of cwMIR3 patches.

We also analyzed misclassifications by the binary classification model. Fig4 shows the percentage of cwMIR1, cwMIR2, and cwMIR3 patches across correct and incorrect diagnoses. Patients with MIR01 that are misclassified as MIR23 higher percentage of cwMIR1, cwMIR2, and cwMIR3 patches. Similarly, MIR23 cases miscalled as MIR01 have a lower percentage of cwMIR1, cwMIR2, and cwMIR3 patches than those with accurate prediction.

Figure 4.

The percentage of cwMIR1, cwMIR2, and cwMIR3 patches in cases of patients with a) no MIR or MIR1 that were correctly predicted or predicted as MIR23, b) MIR2 or MIR3 correctly predicted or predicted as MIR01

White blood cell count and temperature prediction

Table 4 shows performance of UNI-based whole slide regression networks for white blood cell prediction and temperature. Both the models show weak correlation between predicted and actual values.

Table 4.

Performance for white blood cell count (WBC) prediction and fever (Tmax) prediction using UNI

value	Model	RMSE	MAE	R2	slope
WBC	UNI	5.370	4.045	0.071	0.123
Tmax	UNI	1.021	0.856	0.132	0.396

Discussion

Automated analysis of MIR

We investigated the feasibility of predicting MIR stages (MIR0,1 and MIR2,3) from WSI. We found that our attention-based MIL model using UNI foundation model’s features is able to classify MIR with a balanced accuracy of up to 88.4% with a Cohen’s κ of up to 0.758. This is despite the fact that UNI has not been pretrained on any placental data. We also investigated prediction of white blood cell count (WBC), and fever (Tmax) from WSI using two UNI.

Systemic inflammation

Clinical chorioamnionitis is defined by maternal fever, fetal or maternal tachycardia, uterine tenderness, and foul discharge, and is often accompanied by elevated WBC count.^49–51 We found that the regression model with the UNI feature extractor, shows a moderate correlation between the predicted and actual Tmax, indicating a moderate ability to predict fever from histopathological images. Prediction of white blood cell count shows a weak, but still statistically significant correlation. This approach compliments our previous work where we showed neutrophil density correlates with Tmax and WBC.⁴⁷ In contrast to this previous work, the models presented here provide a slide-level gestalt, rather than precise enumeration in an expert-defined region. A future approach, combining region-awareness with cell classification could improve accuracy and interpretability. The interplay between individual features of systemic inflammation and between systemic and histologic inflammation is complex. A more quantitative measure of placental histologic inflammation is part of explaining why histologic and clinical chorioamnionitis are not always co-incident.

Interobserver variability vs misclassification

Interobserver variability among expert pathologists is a significant challenge in placental biology. A predictive model or epidemiologic study using pathologist diagnosed MIR will inevitably suffer if pathologists differ in how they implement the classification scheme. ML models should be consistent, giving the same result every time. In some instances, consistency may be favored over accuracy, for example in the movement from pathologist to automated scoring for certain therapy qualifying immunohistochemical stains in cancer treatment.⁵² We observe that misclassification by our model, are dependent on the frequency of cwMIR1, cwMIR2, and cwMIR3 patches present in the slides. This indicates that the model relies on the amount of target class tissue present in the whole slide. This closely resembles the decision-making process among practitioners. The reliance on the amount of evidence of inflammation is one of the factors that contributes to interobserver variability – that is, different observers will have different thresholds. This highlights the challenges of interobserver variability in building and assessing models. In a different context, we found that training on noisy labels from multiple observers can yield expert level performance.⁵³ Future studies could show similar improvement.

Strengths:

One of the major strengths of this study is that the dataset the large real-world dataset. Further, we have investigated the use of medical imaging foundation model (UNI) as feature extractor in the MIL pipeline, and demonstrated that it achieves a balanced accuracy of 88.4%, for binary MIR classification, even though no placental images were used to train UNI. While, this establishes a baseline of performance for MIR classification, it also highlights the need for finetuning on placenta data for these models to extract finer details from placenta images, and achieve further performance gains. We also investigated UNI’s ability to predict fever from histopathological images, and found moderate correlation between predicted and actual Tmax.

Limitations:

One of the weaknesses of this study is that the dataset was sourced from a single site. The models might not generalize to datasets sources from other sites. Further limitations include that other maternal or fetal signs of inflammation were not considered. The rarity of neonatal sepsis poses challenges to predict it, using our data. Also, due to the nature of our dataset, long-term neonatal outcomes could not be investigated

Figure 1.

(A) Model architecture for MIR stage classification. Each MIR class has a corresponding set of attention weights for the patch embeddings. The aggregated embeddings for each class are processed by fully connected classification layers followed by a softmax layer to convert to class probabilities. The class with highest probability is the model prediction. Attention maps are visualized by coloring each patch based on corresponding attention weight. (B) Model architecture for white blood cell count (WBC) and maternal highest temperature (Tmax) prediction

Data statement

Code and models are available from the corresponding author upon request. Due to medical center restrictions, we are unable to freely distribute WSI. Interested investigators may contact the corresponding author to initiate a data use agreement.