Integrating Clinical Parameters and Cardiotocograms for Real-Time Alerting

Technical Summary: Integrating Clinical Parameters and Cardiotocograms for Real-Time Alerting

Problem Statement
Cardiotocography (CTG) is the standard method for monitoring fetal condition during labor, recording fetal heart rate (FHR), uterine contractions (UC), maternal heart rate (MHR), and fetal movement (FM). However, visual interpretation is subjective, leading to significant inter- and intra-rater variability and inconsistent clinical decisions. While computer-assisted systems exist, they are often limited to retrospective analysis or rigid rule-based logic that cannot discover novel patterns. Furthermore, existing machine learning approaches typically rely on static feature vectors (making them retrospective) or focus exclusively on raw signals without integrating clinical context. There is a lack of systematic evaluation regarding the integration of static clinical parameters (CPs) into real-time, continuous CTG interpretation systems.

Methodology
The study utilizes a retrospective dataset of 254 recordings from 147 patients at Hannover Medical School (January–August 2024). The dataset includes four signals sampled at 4 Hz (FHR, UC, MHR, FM) and static clinical parameters extracted from electronic health records (e.g., maternal age, BMI, obstetric history). Ground truth labels ('Normal', 'Suspicious', 'Pathological') were assigned by two experienced obstetricians based on FIGO guidelines, with a moderate inter-rater agreement (Cohen's $\kappa$ = 0.587).

The authors developed and compared three machine learning architectures, each tested in two configurations: a baseline (CTG data only) and an augmented version (incorporating static CPs). The models were evaluated in a simulated real-time setting, generating predictions at 1 Hz starting from the first minute of available data.

1. Feature-Based Model (XGBoost): An Extreme Gradient Boosting classifier processing engineered features derived from sliding windows (10, 90, and 120 minutes). Features included FHR baseline, variability, event counts, and novel stateful duration features (cumulative time in critical states). Static CPs were concatenated to the feature vector.
2. Raw Data-Based Model (Transformer): A time-series transformer processing raw preprocessed signals (FHR, UC, MHR, and a missing indicator). It utilized a convolutional embedding layer followed by a two-layer encoder. Static CPs were embedded via an MLP and fused per time step.
3. Hybrid Model (LSTM): A Long Short-Term Memory network processing a combined vector of raw signals and engineered features. For the augmented version, an MLP processed static CPs in parallel, creating a clinical embedding concatenated with the LSTM output at every time step.

Key Contributions

• Real-Time Continuous Analysis: Unlike many studies using static vectors or fixed segments, this work implements continuous, second-by-second feature extraction and prediction, addressing the temporal dynamics of fetal state transitions.
• Systematic Evaluation of Clinical Parameters: The study explicitly quantifies the impact of integrating a comprehensive set of static clinical parameters across different architectures, challenging the assumption that more data is always beneficial.
• Hybrid Architecture Superiority: The paper demonstrates that fusing raw signal data with engineered, knowledge-based features yields superior performance compared to purely feature-based or purely raw-data approaches.
• Clinically Oriented Metrics: Performance is prioritized by 'Pathological' recall (sensitivity) and macro F1-score, acknowledging the critical need to minimize false negatives in fetal distress detection while accounting for severe class imbalance.

Results

• Best Performing Model: The hybrid LSTM model without clinical parameters achieved the best overall performance, with a 'Pathological' recall ($R_p$) of 0.565 and a macro F1-score of 0.545. This performance reached approximately 85% of the inter-expert agreement level ($R_p$ = 0.697).
• Impact of Clinical Parameters: The inclusion of CPs was not uniformly beneficial and was highly model-dependent:
  • Transformer: Performance improved with CPs ($R_p$ increased from 0.560 to 0.595).
  • Hybrid LSTM: Performance degraded with CPs ($R_p$ dropped from 0.565 to 0.490), though 'Suspicious' recall improved significantly.
  • XGBoost: Overall performance degraded with CPs.
• Error Analysis: Misclassifications predominantly occurred between adjacent classes (e.g., 'Normal' to 'Suspicious') and at temporal boundaries, reflecting the gradual nature of fetal state transitions and the inherent imprecision of manual annotation.

Significance and Claims
The authors claim that their work lays the methodological groundwork for clinically useful real-time alerting systems. They assert that the hybrid approach is superior for continuous CTG analysis because it leverages the complementary strengths of deep learning (capturing subtle temporal dynamics from raw signals) and feature engineering (integrating high-level clinical logic).

The study modestly concludes that while hybrid models are feasible for automatic alerting, the integration of static clinical parameters requires careful, architecture-specific selection rather than simple inclusion. The authors emphasize that their findings support clinicians in the complex task of CTG interpretation but note that further validation on larger, multi-center datasets and prospective clinical trials are necessary before clinical translation. The paper explicitly states that the goal is not to replicate a single annotator perfectly but to establish a robust system that reduces the subjectivity inherent in current clinical practice.