|
|
Nonlinear Data Analysis and Modelling of Complex Systems:
Prediction of VF Defibrillation Outcomes by Neural
Networks
In this
study, we investigate pattern recognition and classification of ECG
recordings of patients suffering from ventricular bibrillation.
Specifically we develop and assess a nonlinear methodology, using a
combined neural network and statistics approach, to improve the accuracy
and reliability of the outcome predictions.
The VF ECG
data bank was provided by our partner from the Medical Control Module (MCM)
of the defibrillator, Heartstart 3000 (Laerdal Medical, Stavanger,
Norway) and the regular
Utstein registration in Ulleval
University Hospital,
Oslo, Norway. In this study, 150
pre-shock episodes of VF from 59 cardiac arrest patients with
defibrillation conversions to different outcomes were selected on the basis
of low noise and lack of artefacts. The ECG data were sampled at 100 Hz
with 8-bit A/D resolution. The typical length of a VF pre-shock ECG episode
is 2000 sampling points in 20 seconds. Among the total 150 pre-shock VF ECG
episodes, we use two sets of 75 episodes corresponding to post-shock
conversions to ROSC and non-ROSC. Within the 75 ROSC cases, there are 30
episodes of sustained ROSC and 45 of non-sustained ROSC.
We
investigated two different strategies for predictions of outcomes of
electric defibrillation using pre-shock VF ECGs. Both investigations were
based on probabilistic neural networks (PNNs) but used different procedures
for obtaining input patterns; the first simply using the raw ECG while the
second used features extracted statistically from the raw data. We chose
the robust heteroscedastic probabilistic neural network (RHPNN) since it
successfully deals with uncertainties in decision making by combining
empirical experience and prior knowledge to minimise the probability of
misclassification and has the advantages of parsimonious structure and
robustness in dealing with noisy data. In the first study, the features (input
patterns to the RHPNN) were organised through delayed co-ordinates, based
on embedding theory, of the VF ECGs. The trained RPHNN was shown to
have the ability to cluster the input patterns into two classes
corresponding to patients with or without ROSC after a defibrillation
shock. The classification capability of the RHPNN classifier was evaluated
by its test performance measured through sensitivity and specificity. Our
experiments show that this method has achieved 74.5% in overall
classification accuracy with sensitivity 84% and specificity 65%. The
classification can be made with short VF ECG segments of an order of one
second. In the second study, the features used for the outcome
prediction were 6 statistic measurements – 4 based on the power spectrum
density of the VF ECGs and computed for three different frequencies domains
(0 to 5Hz, 12.5 Hz and 25Hz), and the other 2 corresponding to the average
period and amplitude of the ECGs. Importantly, we show that this predictor
is capable of performing multiple-stage predictions. For the first stage,
ROSC vs non-ROSC, we obtained further improved results over those of all
previous studies; with sensitivity 93.6% and specificity being 74.0%. At
stage two, sustained ROSC vs non-sustained ROSC, sensitivity of 94% and
specificity of 82.0% are achieved. We further attempted the third stage
predictions, survival and non-survival, and obtained comparable prediction
accuracy. However, due to the limited number of survival patients in the
data bank (only 10 cases), the reliability of the prediction needs to be
further examined. These results have convincingly shown that features from
VF ECG time series prior to defibrillation shocks can be extracted and used
as the inputs to neural networks for effective classification and
prediction of their post-shock results
Back to Research Reviews'
Page
|
|
