A machine-learning based analysis for the recognition of progressive central hypovolemia

Healthy, non-smoking subjects aged between 18 and 50 years were recruited by advertisement in the Academic Medical Centre in Amsterdam (AMC), excluding any subject with a history of and/or treatment for cardiovascular disease, diabetes, and fainting. The measurements were performed between 6 January 2015 and 11 June 2015. Prior to inclusion, a full medical history was obtained, followed by physical examination and oscillometric BP as well as ECG measurements to assess the presence of cardiovascular contra-indications for the procedure. The study was approved by the Medical Ethics Committee of the AMC. Written informed consent was obtained from all subjects before any measurement.

Measurements were performed in a quiet room. Subjects were informed about the procedures involved and instructed to empty their bladder prior to the start of testing. They abstained from coffee, tea, sports, and fatty foods the day before measurement. With the subject in the supine resting position a lower-body negative pressure (LBNP) box was placed over the lower body. The LBNP box sealed the subject from hip to toes, enabling the buildup of a pressure difference between the upper and lower body. It was verified whether the box created and maintained a subatmospheric pressure of  −50 mmHg. A baseline of 5 min, consisting of the same parameters obtained during LBNP, was measured after 25 min of supine rest. A LBNP of  −50 mmHg was induced after the baseline and maintained for 30 min or terminated earlier in the case of the development of signs or symptoms of pre-syncope such as lightheadedness, nausea, or blurry vision. Pre-syncope was additionally defined as a sudden consistent reduction of  >25 mmHg in systolic blood pressure (P_sys), of  >15 mmHg in diastolic blood pressure (P_dia), or of  >15 beats min⁻¹ in HR, as these parameters will change with a large (>30%) decrease of CBV (Secher and van Lieshout 2016). The start and end of the baseline and LBNP was indicated by markers. Data from the subjects not experiencing pre-syncope within 30 min were excluded from the analysis.

BP was measured by plethysmography at the middle finger of the left hand (Nexfin, Edwards Lifesciences BMEYE, Amsterdam, The Netherlands) during the entire experiment. The signal was stored at 200 Hz. TI (Model AI-601G, Nihon Kohden, Japan) was measured by placing electrodes in the neck and on the upper body at the level of the xiphoid process. TI measures conduction through the chest as a gauge of CBV. The analog TI signal and a marker device were sampled at 200 Hz and stored on disk. The raw BP curve was extracted from the Nexfin. All signals were resampled at 100 Hz using Matlab (2016A, The MathWorks, Natick, MA, USA). Nine beat-to-beat parameters were extracted directly from the Nexfin. These parameters consist of P_sys, P_dia, mean arterial pressure (P_mean), P_pulse, HR, SV, CO, SVR, and left ventricular ejection time (LVET). Corresponding beat-to-beat values of TI were added as a parameter. Age, gender, height, and weight were entered as static parameters.

Analysis was performed off-line using a custom-made Matlab algorithm combined with the neural network toolbox in Matlab. The LBNP period was selected based on the manually defined markers. A 30 s baseline period containing few or no artifacts was selected out of the recorded baseline to accurately reflect the resting state of the subject. Recordings during the first 120 s of LBNP were removed to discard the sudden hemodynamic changes that occur at the start of LBNP. Data was visually inspected for artifacts in the raw BP curve and in the beat-to-beat parameters HR, SV, SVR, and TI. Artifacts were removed with subsequent interpolation of the remaining data.

All beat-to-beat parameters were normalized to their median value within the selected baseline period. Thereafter, these were smoothed using an averaging filter with a moving 30 s window. Because time to pre-syncope and HR varied between subjects, the amount of beat-to-beat parameter values estimated for each subject differed between subjects. To ensure that each subject contributed an equal number of samples to the model, subject data were resampled to the median number of beat-to-beat parameter values over all subjects.

Model design and statistics.

The input to the neural network for the prediction at the time point of each heartbeat contained ten beat-to-beat parameters measured at that single heartbeat, as well as static data on age, sex, height, and weight (figure 1). Time to pre-syncope was scaled from 100%–0%, with 100% defined as 120 s after initiation of LBNP and 0% as the onset of pre-syncope. Time to pre-syncope was the desired output of the model for each training instance.

Zoom In Zoom Out Reset image size

Data was divided into a training set and test set. The training set consisted of all but one subject. The test set consisted of the remaining subject. Each subject served once as a test set. For each test set, 100 neural networks were trained using the training set, as more neural networks did not increase performance, while it did increase computation time. All neural networks were initialized randomly. Thereafter, models were trained using the Levenberg–Marquardt algorithm for backpropagation. Each of the 100 neural networks predicted the time to pre-syncope for the selected test subject. These 100 predictions were averaged to create a single average prediction of the selected test subject. We assumed normovolemia at the initiation of LBNP and therefore the average predicted time to pre-syncope was set equal to 100% at the start. Each of the tested subjects served once as a test set (leave-one-out cross-validation) in the above-described procedure, finally resulting in an average prediction for each subject. These predictions ideally start at 100% when the true time remaining is also 100%, and indicate 0% when the true time remaining is also 0%. Therefore, when plotting the true time remaining versus the predicted time remaining, the ideal slope would be  −1.

The above-mentioned model was created with a variable number of nodes in the hidden layer and a variable number of iterations during the training phase (figure 1), influencing the complexity of the model and the amount of training on the data. The number of hidden nodes were 5, 10, 15, 20, or 25 (figure 1(a)) and the number of iterations 10, 20, 40, 80, 160, or 320 (figure 1(b)). This resulted in a total of 30 combinations of hidden neurons and training iterations. Because each combination leads to an average prediction over 100 neural networks, the result was a grid of average predictions for each subject. A straight line was fitted to each of the 30 average predictions, the slope of which should be ideally  −1, i.e. equal to the expected target slope. For each combination of hidden nodes and number of iterations, the slopes were subsequently averaged over all subjects. Each averaged slope can be interpreted as the expected slope of the prediction on a new subject. The least complex combination of hidden nodes and iterations that led to a slope closest to  −1 with the lowest standard deviation (SD) was considered the optimal model. Using the outcome on the test set to select the optimal model may introduce an optimistic outlook. However, no contamination between the training and test set is expected, because the data are already trained and tested at the moment of optimal model selection, thus reducing the possibility of an overly optimistic outlook. The fitted line on a prediction is used to evaluate the quality of the neural network, while the actual predictions by the model assess the quality of the model-prediction in an individual subject.

Machine learning techniques were also used to develop the compensatory reserve index (CRI), indicating the amount of reserve before pre-syncope will occur in a subject on a scale from one to zero (Convertino et al 2011, Moulton et al 2013). The CRI is shown to be more predictive for hypovolemia than the commonly used shock index, which is the ratio of HR to P_sys (van Sickle et al 2013). Furthermore, a decrease in CRI is seen when a small amount of blood (450 ml) is withdrawn from a subject (Nadler et al 2014, Stewart et al 2014). However, when withdrawing an average of 1.2 l in 20 subjects, the CRI decreased on average by 33%. This small decrease in CRI suggests that the CRI underestimates larger amounts of blood loss (Convertino et al 2015).

An important difference between our model and the CRI is the collected data on which to train the respective models. The CRI model is trained on subjects who underwent a stepwise LBNP protocol in which LBNP is lowered after every 5 min by 5 mmHg, until pre-syncope occurs. The CRI is based on the prediction of the current level of LBNP and the predicted level of LBNP at which pre-syncope will occur. During a 5min period at a certain level of LBNP, both the predicted current and final level of LBNP will be constant if the algorithm is correct. Therefore, during this 5 min period, the CRI will be stable. This is also seen in the outcome, as the predicted levels of LBNP show a stepwise fashion, with high correlation to the actual level of LBNP (r  =  0.94) (Convertino et al 2011). When estimating the CRI out of these levels of LBNP, the result is a stepwise CRI (Moulton et al 2013). However, even though LBNP is constant between two steps, there is still a change in CBV, which is not taken into account by the CRI. In our study, a constant level of LBNP was used, resulting in a continuous change in CBV during the procedure. Therefore, our model is trained to detect a change in CBV without the effect of the change in LBNP, which may be closer to the real-life situation of central hypovolemia. Furthermore, our study is to be applied in the operating room (OR). As the CRI is developed for the battlefield, our model can take more parameters into account. The difference in the model as well as the target group makes this a good addition for use in the OR.

First, the number of subjects in our neural network is relatively small. Although this drawback is partially compensated for by the low number of parameters, increasing the number of subjects may increase the accuracy of the prediction. Furthermore, increasing the number of subjects opens the possibility of creating more diversity in the models by subset selection of the subjects, leading to possible better predictions. Due to the low number of subjects, no validation set was used in this study. This may introduce an overly optimistic outlook for the selected optimal model. In this study, however, a clear downward trend was found for all possible combinations of hidden neurons and iterations (table 1). This suggests that a good prediction model would have been found for all possible combinations tested. For future studies, the inclusion of a validation set together with more subjects is recommended. Second, measurements were discontinued if the vasovagal response developed or if a subject indicated symptoms of pre-syncope. Because the onset of symptom occurrence is not by definition equal to the moment of pre-syncope, some measurements might have been interrupted early. Although this fact warrants subgroup analysis based on stopping criteria, this cannot be performed due to the low number of subjects. Third, we set the average prediction at the start equal to 100%, because we assumed normovolemia at the initiation of LBNP. It is realistic that the subjects were slightly hypovolemic, setting the true initial value below 100%. This decreases the slope of the reference line and may indicate that the calculated slope for the subjects may be closer to the reference slope than stated previously. Finally, the model requires measurement of a baseline period, which may not be available in acute cases. In these cases, baseline values may be estimated based on patient characteristics. A different approach would be to abandon normalization of parameters to the baseline, though possibly at the expense of prediction accuracy. The possibly lower prediction accuracy when using non-normalized parameters might be counteracted by the inclusion of trend parameters.