Laparoscopy is a minimally invasive surgical intervention that requires high proficiency skills such as depth perception and hand-eye coordination. In addition, surgeons must develop cognitive skills such as attention, information processing, storage and feedback of information, in order to make the best decision at the right time, and execute the necessary movement for the patient well-being [28]. These skills are affected by lack of rest (a common practice in the medical field). For this reason, this study aimed to identify the states of vigil and fatigue due to lack of sleep, through the analysis of EEG activity. Fatigue state in surgeons can lead to execute irreparable surgical errors in patients. Hence, there is an opportunity to eventually implement systems that alert about the current state of the surgeon, and thus prevent accidents.
EEG activity is sectioned into various rhythms or frequency bands: gamma (> 30 Hz), beta (13–30 Hz), alpha (8–13 Hz), theta (4–8 Hz) and delta (1–4 Hz) [29]. In particular, theta and alpha bands are related to learning processes. Theta band has been associated with memory and cognitive tasks; synchronization of this rhythm is due to the increase of the cognitive effort during the execution of an activity. In general, this band is present over the frontal lobe, which is associated with visual perception [30, 28]. On the other hand, alpha band is related to the level of attention, thus desynchronization of this rhythm is related to inhibitory action to enable the processing of information [28]. On this evidence, the present study seeks to analyze the amplitude changes of EEG signals in theta and alpha bands during five laparoscopic training tasks of 18 junior surgeons, before and after their duties to differentiate between vigil and fatigue states. It was found that the required time to perform the laparoscopic tasks did not change significantly before (vigil) and after (fatigue) the on-call (statistically speaking). However, the central trends in execution times were greater or equal after the duty, except for the ligating loop activity (Fig. 5). This means that on average, junior surgeons in training spent more time to perform the same task due to lack of sleep [31].
On the other hand, the power changes of the spectra of the EEG signals in theta and alpha bands were enough to differentiate between wakefulness and fatigue states (accuracy = 90.7%). That is, 9 out of 10 vectors of EEG features recorded in real time could be properly identified within the correct state: wakefulness or fatigue. The results of the present study contribute with a first attempt to identify the mental states of physicians in training under real conditions. It should be noted that the experimental procedure was not implemented in a laboratory under controlled conditions, but in the Skills Center, where physicians are trained to master the technique of laparoscopy. Furthermore, the presented results can be improved with other analysis strategies. For example, in [32], authors classified 40 different signals from each electromagnetic disturbance (320 in total), which interfere with the normal operation of electrical power systems by using the Hilbert-Huang Transform with an accuracy of 94.6%. Another improvement can be the use of neural networks as a classification phase. As a preliminary analysis, three different neural networks were used:
-
InceptionV3 Model implemented in Python using Keras applications. The architecture can be consulted in [33], its weights were loaded from ImageNet to have a pre-trained network, and the signals were converted to standardized images, and rescaled to have a size of 299x299x3.
-
Model that concatenates the feature vectors of the VGG19 and ResNet50 models. The model architecture can be seen in [34], only AlexNet was replaced by ResNet50 as it was the one available in Keras applications. Keras applications was used for both models, and the weights were imported from ImageNet. Each image was normalized and rescaled to 227x227x3 to feed data in the concatenated model.
-
5-Fully-Connected layer Neural Network [35]. The first layer had 500 units, the second one had 290 units, the third one had 79 units, the fourth one had 59 units, and a single output. The network iterated 10000 times, and the learning rate parameter was 0.0075.
Using the same characteristics, the same number of observations, and the same distribution of data (70% for training and 30% for testing), the performance achieved of the three neural networks used in the preliminary analysis did not exceed 73%, as shown in Table 4.
Table 4
Classification of three neural networks
Neural network | Accuracy |
Inception V3 | 66.33% |
VGG19ResNet50 | 50.57% |
5-Fully-Connected | 73.08% |
The presented accuracies are low, practically the networks are guessing if the junior surgeon is in a state of vigil or fatigue, except for the 5-layer neural network totally connected that showed the best result. This is because the first two networks detected a high variance in data. To improve the results, it is necessary to increase the number of observations and/or include regularization parameters to the network, which implies a higher computing time. The last network could be optimized by adding more layers, or by training the network for a longer period, what would become unfeasible for the purposes of this research.
The findings of the present research broadly support the work of other studies in this research area. In [36], authors classified manual movements (left and right) of experienced and novice residents, after the use of an obstetric clamp. They obtained a classification of 99.1% with the employment of a recurrent neural network (RNN), along with the long-term-short-term memory (LSTM). In [37], authors were able to classify EEG signals between wakefulness and fatigue states, reaching an accuracy of 96.28%. The methodology proposed by [38] achieved 99.61%. In [39], it was found that stress (analyzed with alpha band records) is a factor that negatively influences the execution of surgical practices. In [40], authors analyzed fatigue in surgical practices, by repeating the peg transfer task. They concluded that effectiveness in surgical practice is not determined by the time taken to perform an operation, but by the absence of errors. These studies lead us to propose that the execution time of the laparoscopic tasks reported in this study (Table 2) can be eventually used to correlate the effectiveness and execution time of each task, affected by states of wakefulness or fatigue.
In conclusion, sleep-induced fatigue generates power changes of identifiable EEG signals in theta (front-center area) and alpha (temporal and parieto-occipital areas) bands in trainee physicians during their laparoscopic training. The EEG analysis was based on differentiation between absolute power values through a cubic SVM classifier.