prediction of intrauterine pressure from electrohysterography using optimal linear filtering mark d....
TRANSCRIPT
Prediction of intrauterine pressure from electrohysterography using optimal
linear filtering
Mark D. Skowronski
Computational Neuro-Engineering Lab
Electrical and Computer Engineering
University of Florida
Gainesville, FL, USA
August 31, 2005
Overview• Introduction
• What are IUP and EHG?
• Previous studies
• Wiener filter prediction
• Results and discussion
• Conclusions and future work
Collaborators
• Neil Euliano* (P.I.), Convergent Eng., Gainesville, FL• John Harris*, Assoc. Prof. ECE, CNEL, UF• Tammy Euliano, Assoc. Prof. Anesthesiology, UF• Dorothee Marossero*, Convergent Eng., Gainesville, FL• Rod Edwards, Obstetrics and Gynecology, UF• Support from NSF, DMI-0239060
* = current/former members of CNEL
Introduction• Biology inspires models
– Human factor cepstral coeffs– Energy redistribution– Freeman model, ESN, LSM– Spike-based circuits, algorithms
• Apps. with biological signals– HFCC, ER– Bat acoustics– Brain-machine interfaces– EEG, fMRI research– Electrohysterography
BIOLOGYMODELS
Prenatal monitoring• Intrauterine pressure (IUP)• Tocodynamometry (Toco)• Electrohysterography (EHG)• Ultrasound
Labor monitoring• Intrauterine pressure
– Uterine muscle activity (contractions) exerts force on the fetus towards cervix.
– Force is measured using intrauterine pressure catheter (IUPC).
– Used to monitor progression of labor.
• IUPC limitations– Used only after membrane rupture.– Internal, invasive technique, infection risk.– Requires presence of obstetric indicators to
justify risk.
Labor monitoring• Electrohysterography
– Skin electrodes, noninvasive.– Macroscopic muscle activity.– Multiple simultaneous measurements possible, more
information about labor state.– Useful throughout pregnancy.
• EHG limitations– Difficult to reliably measure muscle activity through
skin.• Variable skin resistance, preparation.• Variable distance to muscles (fetal shifts).
– Electrode placement repeatability.– Indirect monitoring method.
EHG and IUP example
Previous EHG studies• Correlation with IUP
– Generated from same underlying phenomenon.– Hand-excised contractions, correlation
• IUP feature: integral• EHG feature: energy between 0.3-1.0 Hz• r = 0.76, Maul et al., 2004
• Predicting delivery– EHG feature: spectral peak freq., 0.3-1.0 Hz– Peak freq. increases as time to delivery decreases– Accurate 24 hours before delivery, Maner et al., 2003
• No previous studies of continuous IUP prediction from EHG
IUP prediction from EHG• Proposed method: Wiener filter solution– y(n)--model output– x(n)--EHG input– w(n)--Wiener filter coefficients, length N
• Properties– Causal, linear FIR filter, optimal in MSE sense.– Closed-form solution, easy to train.– Output is projection of input space onto vector of
filter coefficients, real-time implementation.– Competent baseline algorithm, useful in developing
future more sophisticated prediction models.
1
0
)()()(N
i
inxiwny
Methods• Data collection
– 303 pregant females monitored at Shands between July 2003 and Jan. 2005.
– 8-channel EHG data was collected, 200 samples/sec/channel, 16-bit resolution.
– Of those, 32 simultaneously monitored with IUPC, 2 samples/sec, 8-bit resolution.
– Of those, 14 remained after screening• At least 30 minutes of data (10 patients)• At term (3 patients)• No obvious data artifacts (5 patients)
Methods, con’t• IUP signal preprocessing
– Non-causal median filter, ±5 seconds, to remove spiky noise.
– Downsampled from 2 Hz to 0.2 Hz
Methods, con’t• EHG signal preprocessing
1. Zero mean, unity variance.2. Downsampled from 200 Hz to 4 Hz
(relavent bandwidth from literature).3. Rectified (nonlinear operation, crude
energy estimate).4. Downsampled from 4 Hz to 0.2 Hz (shorter
filters, faster training, no affect on under training).
Experiments• Single channel, single patient
– 10-minute test/train windows– Each line below is from the best model/best
channel/best test window for each patient (test-on-train results excluded)
Performance saturates at 50 sec.
Experiments, N = 50 sec• Single channel, single patient
– Each group of points is from the best model/best test window for each patient/channel
Prediction examples, N = 50 sec
Pt. 41, ch. 2, r = 0.90, RMS error = 3.7 mmHg
Pt. 229, ch. 8, r = 0.86, RMS error = 10.0 mmHg
Analysis of variance• 4-way ANOVA
– Dependent variable: RMS error.– Independent variables: patient, channel, time (test
window), model (train window).– All interactions not listed below were insignificant.
Factor d.f. F p Range, mmHg
Patient 13 21.8 0 5.2-13.7
Channel 7 0.76 0.62 9.3-10.3
Time 16 30.3 0 8.7-11.2
Model 16 11.2 0 9.2-10.5
Pt*Ch 91 16.9 0 3.4-17.6
Ch*Time 112 3.4 0 7.3-12.1
Ch*Model 112 0.76 0.98 8.7-11.8
Conclusions• Wiener filter/rectified EHG useful for
predicting IUP– Best of the best: r > 0.90, RMS error < 9
mmHg– RMS error sensitive to factors: patient, time,
model, pt*ch, ch*time, ch*model– RMS error not sensitive to factors: channel,
pt*time, pt*model, time*model, all higher interactions
Future work• Better figures of merit
• Single patient, multi-channel
• Multi-patient, multi-channel
• Better features besides rectified EHG
• Non-causal Wiener filter
• More powerful prediction models
• Weighted RMS error/squared prediction