Sparse learning for intrapartum fetal heart rate analysis

The paper provides results on Lyon database and selected cases from CTU-UHB databases. The reasoning for CTU-UHB selection is provided in the paper, full results are presented below CTU-UHB database - results on 552 cases.

P. Abry, J. Spilka, R. Leonarduzzi, V. Chudáček, N. Pustelnik, M. Doret Sparse learning for Intrapartum fetal heart rate analysis In Biomedical Physics Engineering Express 4(3) 034002, 2018.

Abstract

Fetal Heart Rate (FHR) monitoring is used during delivery for fetal well-being assessment. Classically based on the visual evaluation of FIGO criteria, FHR characterization remains a challenging task that continuously receives intensive research efforts. Intrapartum FHR analysis is further complicated by the two different stages of labor (dilation and active pushing). Research works aimed at devising automated acidosis prediction procedures are either based on designing new advanced signal processing analyses or on efficiently combining a large number of features proposed in the literature. Such multi-feature procedures either rely on a prior feature selection step or end up with decision rules involving long lists of features. This many-feature outcome rule does not permit to easily interpret the decision and is hence not well suited for clinical practice. Machine-learning-based decision-rule assessment is often impaired by the use of different, proprietary and small databases, preventing meaningful comparisons of results reported in the literature. Here, sparse learning is promoted as a way to perform jointly feature selection and acidosis prediction, hence producing an optimal decision rule based on as few features as possible. Making use of a set of 20 features (gathering ‘FIGO-like’ features, classical spectral features and recently proposed scale-free features), applied to two large-size (respectively ;1800 and ;500 subjects), well-documented databases, collected independently in French and Czech hospitals, the benefits of sparse learning are quantified in terms of: (i) accounting for class imbalance (few acidotic subjects), (ii) producing simple and interpretable decision rules, (iii) evidences for differences between the temporal dynamics of active pushing and dilation stages, and (iv) of validity/generalizability of decision rules learned on one database and applied to the other one.

Database

Two independent large-size databases are used. They were collected—with different technologies and constraints—at academic hospitals in Lyon, France (LDB), and Brno, Czech Republic (CTU-UHB). They share comparable clinical characteristics. Tracings were included according to clinical and data-quality criteria (Doret et al. 2011, Spilka et al. 2017): gestational age >= 37 weeks, maternal age >= 18, tracing ending less than 20 minutes before delivery, after-delivery pH measurement available, less than 50% of missing data in either labour stage

CTU-UHB database - results on 552 cases

In the paper the subjects were selected based on the data quality at the end of first stage and second stage: i) tracing ending less than 20 minutes before delivery, ii) less than 50% of missing data in either labour stage.

The goal was to explore the benefits of sparse learning in the construction of decision rules. The actual computation of usable features required a decent data quality. Therefore, to avoid the mixing of two independent issues : i) benefits of sparse learning ii) impacts of the quality of the data, we have decided to focus on the first issue mostly avoiding the blurring of conclusions on benefits of sparse learning by data quality issues. However, to ease comparison of future works we present here results on the full CTU-UHB database (552 subjects).

To compute the performance on the full database we have selected the last 20 minutes with FHR quality above 50%. We used the stage I model (SI) for records ending in the first stage of labour and stage II model (SII) for records ending in the second stage.

Results:

stage:	SE	SP	BER	TP	FN	TN	FP
SI	0.62	0.81	0.72	20	12	317	75
SII	0.58	0.77	0.68	7	5	89	27
sum SI + SII	0.61	0.80	0.71	27	17	406	102

SE - sensitivity, SP - specificity, BER - balanced error rate, TP - true positive, FN - false negative, TN - true negative, FP - false positive

Detailed prediction for each subject:

Prediction of sparse SVM for each subject in a csv format: Abry2018_BPEX_CTU_UHB_results.csv