Antenatal surveillance of placental function using a wearable near infrared spectroscopy device with machine learning data interpretation

This study demonstrates that a wearable near-infrared spectroscopy device combined with machine learning analysis can accurately predict adverse pregnancy outcomes by utilizing dynamic metabolic and hemodynamic signals, rather than static oxygenation levels, to monitor placental function in high-risk pregnancies.

The Gist

Imagine a pregnancy as a high-stakes delivery of a precious package (the baby) from a factory (the mother) to a destination. The placenta is the factory's shipping dock and power plant combined. It's responsible for delivering oxygen and nutrients while removing waste. If this dock starts failing, the baby can get sick or, in the worst cases, the delivery fails entirely.

Currently, doctors have to check on this "shipping dock" using tools that are like taking snapshots. They use ultrasound or listen to the baby's heart rate, but these only give a quick picture of what's happening right now. They miss the long-term trends and the subtle warning signs that happen in between checks.

This paper introduces a new invention called FetalSenseM (FSM v1), which acts like a continuous, wearable smartwatch for the placenta.

Here is a simple breakdown of how it works and what the researchers found:

1. The Device: A "Smart Flashlight" for the Belly

The team built a wearable device that uses Near-Infrared Spectroscopy (NIRS). Think of this as a special flashlight that shines invisible light through the mother's belly.

• How it works: Just like how a prism splits white light into a rainbow, this device shines light of different colors (wavelengths) into the tissue.
• What it sees: The light bounces back, and the device analyzes the "echo." It can tell the difference between oxygen-rich blood, oxygen-poor blood, and even the battery level of the placenta's cells (a molecule called cytochrome-c-oxidase or oxCCO).
• The Upgrade: Previous versions of this technology were huge, clunky machines that could only work on the front of the belly. This new version is a portable prototype that can work even if the placenta is at the back of the uterus.

2. The Problem: The "Snapshot" vs. The "Movie"

The researchers tried to use this device to see if the placenta's oxygen level (how much fuel is in the tank) could predict a bad outcome.

• The Result: Surprisingly, just looking at the average oxygen level didn't work. It was like trying to predict a car crash by only looking at the average speed of the car over the last hour. Sometimes the car was speeding, sometimes slow, but the average looked fine.
• The Twist: They found that in some difficult pregnancies (like severe growth restriction), the placenta actually had higher oxygen levels. Why? Because the "shipping dock" was so clogged that the oxygen couldn't get to the baby, so it just piled up at the dock. This is a paradox: high oxygen at the dock, but a starving baby.

3. The Solution: The "AI Detective" (Machine Learning)

Since the simple "oxygen level" number wasn't enough, the researchers brought in a Machine Learning (ML) AI.

• The Analogy: Imagine a detective trying to solve a crime. Looking at just one clue (the oxygen level) isn't enough. But if you look at the pattern of clues—how the heart rate changes, how the blood flow pulses, and how the cellular "battery" fluctuates over time—the pattern becomes obvious.
• The Magic: The AI analyzed the dynamic signals (the movie, not the snapshot). It looked at how the blood flow and the cellular energy usage danced together.
• The Result: The AI became a very good detective. It could predict adverse pregnancy outcomes (like stillbirth risk) with 78% accuracy.
  • It realized that the relationship between blood flow and cellular energy (the "coupling") was the key. When these two stop dancing in sync, it's a major red flag.

4. Key Takeaways in Plain English

• Static numbers are boring: Just knowing the placenta has "50% oxygen" doesn't tell you if the baby is in trouble.
• Patterns are powerful: The way the oxygen and energy levels change over time tells the real story.
• Metabolism matters: The device found that the "battery health" of the placenta cells (metabolism) is a better warning sign than just the oxygen level.
• It works on everyone: The device worked on women with placentas in different positions (front, back, top), which previous tools couldn't do.

The Bottom Line

This study is like the prototype of a new generation of pregnancy monitoring. It suggests that instead of just taking a quick photo of the baby's heart rate, we should be watching a continuous movie of the placenta's health using a wearable sensor and an AI brain.

While the device still needs to be made smaller and fully wireless (it's currently connected to a laptop by a wire), this research proves that we can "listen" to the placenta's metabolic heartbeat and use that data to catch problems before they become emergencies. It's a shift from reacting to crises to predicting them before they happen.

Technical Summary

1. Problem Statement

Placental dysfunction is a leading cause of stillbirth and neonatal morbidity, yet current antenatal surveillance tools (e.g., ultrasound Doppler, cardiotocography/CTG) are limited. They provide only intermittent, indirect, or surrogate measures of fetoplacental wellbeing, often failing to detect critical hypoxia until it is too late or causing unnecessary interventions due to poor sensitivity. There is a critical need for a non-invasive, continuous monitoring system capable of assessing both placental oxygenation and metabolic function in real-time.

2. Methodology

A. Device Development: FetalSenseM v1 (FSM v1)

• Technology: The authors developed a wearable, continuous-wave (CW) near-infrared spectroscopy (NIRS) device. Unlike previous research-grade time-domain systems, this is a portable prototype designed for clinical use.
• Hardware: It features two light-emitting diode (LED) sources and two photodiode detectors with source-detector separations of 3 cm and 5 cm.
• Wavelengths: It emits light at five discrete wavelengths (780, 810, 830, 850, and 890 nm) to target hemoglobin (HbO2, HHb) and the mitochondrial enzyme cytochrome-c-oxidase (oxCCO).
• Signal Processing:
  • Oxygenation: Uses Spatially Resolved Spectroscopy (SRS) and Dual Slope (DS) analysis to derive absolute placental oxygen saturation (PltO2).
  • Metabolism: Uses a modified Beer-Lambert Law (UCLn algorithm) to resolve relative concentration changes in oxCCO.
  • Validation: Monte Carlo simulations and optical phantom studies were used to validate depth sensitivity, confirming the 5 cm channel effectively targets the placenta while the 3 cm channel targets superficial tissues.

B. Study Design

• Cohort: A prospective observational study at University College London Hospital (UCLH) involving 58 high-risk pregnant women (singleton pregnancies with risks like pre-eclampsia, diabetes, or fetal growth restriction).
• Protocol: Participants underwent ~40 minutes of transabdominal monitoring. Placental depth and skin tone were measured to calculate a Minimum Placental Sensitivity (MPS) threshold. Only recordings with >5% placental signal contribution were included in a refined sub-cohort (33 sessions from 30 participants).
• Outcome Classification: Adverse outcomes were defined using the "In Utero near-miss criteria for stillbirth" (including stillbirth, severe pre-eclampsia, severe FGR, neonatal CPR, etc.).

C. Data Analysis & Machine Learning (ML)

• Feature Engineering: 328 features were extracted, including static indices (mean PltO2) and dynamic features (frequency-domain, wavelet-derived measures).
• Wavelet Analysis: Used to calculate HbD:oxCCO semblance, a metric quantifying the phase relationship (coupling) between hemodynamic oxygen delivery and mitochondrial metabolic function.
• ML Pipeline:
  • Preprocessing included DBSCAN for outlier removal and imputation for missing values.
  • 11 Classifiers were tested (including SVM, Random Forest, XGBoost, Logistic Regression).
  • Validation: Nested stratified 5 × 4 cross-validation (5 outer folds for testing, 4 inner folds for hyperparameter tuning via Optuna).
  • Feature Selection: Top-k feature subsets (ranging from 5 to 100) were evaluated to identify the most predictive variables.

3. Key Contributions

1. First Wearable CW-NIRS for Placenta: This is the first wearable device capable of simultaneously monitoring placental oxygenation and mitochondrial metabolism in vivo across diverse placental positions (including posterior).
2. Novel Application of ML: This is the first study to apply machine learning to placental NIRS signals, moving beyond simple univariate comparisons to complex pattern recognition.
3. Metabolic Biomarkers: The study identifies that dynamic metabolic signals (specifically oxCCO and its coupling with hemodynamics) are superior predictors of adverse outcomes compared to static oxygenation levels.

4. Results

Physiological Findings

• Static Oxygenation: Mean PltO2 was approximately 49.8% across the cohort. Static PltO2 values did not significantly differ between pregnancies with good vs. adverse outcomes in the full cohort.
• Sub-cohort Specifics (MPS >5%):
  • Severe Fetal Growth Restriction (FGR): Showed significantly higher PltO2 (59.79% vs 50.26%, p=0.04). This suggests impaired oxygen extraction (oxygen accumulation in the intervillous space due to poor transfer to the fetus).
  • Gestational Diabetes (GDM): Showed significantly lower PltO2 (49.61% vs 53.73%, p=0.04), likely indicating increased oxygen consumption/metabolic demand.
  • Hemodynamic-Metabolic Coupling: Severe FGR cases exhibited significantly increased HbD:oxCCO semblance (p=0.0002), indicating a shift toward passive, deranged metabolic autoregulation.

Machine Learning Performance

• Best Model: Support Vector Machine (SVM) using the top 50 features.
• Performance Metrics (Full Cohort):
  • Balanced Accuracy: 78%
  • Recall (Sensitivity): 72%
  • Specificity: 84%
• Sub-cohort Performance: The "Extra Trees" model with top 20 features achieved 80% for all metrics (accuracy, recall, specificity).
• Feature Importance:
  • Dominant Predictors: Metabolic features (oxCCO-derived) and hemodynamic-metabolic coupling (semblance) were the top predictors.
  • Non-Discriminatory: Static PltO2 values had negligible predictive power, confirming that univariate analysis is insufficient.

5. Significance and Conclusion

• Paradigm Shift: The study demonstrates that the value of placental NIRS lies not in static oxygenation snapshots, but in dynamic metabolic monitoring. The coupling between oxygen delivery and mitochondrial utilization is a critical biomarker for placental insufficiency.
• Clinical Potential: The integration of wearable NIRS with advanced ML analytics offers a promising tool for continuous, non-invasive antenatal surveillance, potentially reducing stillbirth rates by detecting metabolic distress earlier than current methods.
• Future Directions: While the prototype (FSM v1) requires a wired connection, the findings validate the concept. Future work must focus on miniaturization, wireless transmission, and validation in larger, multi-center, and multi-ethnic populations to refine the predictive algorithms for routine clinical use.

Conclusion: The study successfully proves that a wearable NIRS device can capture complex placental physiology, and when analyzed via machine learning, can accurately predict adverse pregnancy outcomes based on metabolic dysfunction rather than simple oxygenation levels.