Artificial Intelligence and Machine Learning in Anesthesiology

Abstract

Commercial applications of artificial intelligence and machine learning have made remarkable progress recently, particularly in areas such as image recognition, natural speech processing, language translation, textual analysis, and self-learning. Progress had historically languished in these areas, such that these skills had come to seem ineffably bound to intelligence. However, these commercial advances have performed best at single-task applications in which imperfect outputs and occasional frank errors can be tolerated.The practice of anesthesiology is different. It embodies a requirement for high reliability, and a pressured cycle of interpretation, physical action, and response rather than any single cognitive act. This review covers the basics of what is meant by artificial intelligence and machine learning for the practicing anesthesiologist, describing how decision-making behaviors can emerge from simple equations. Relevant clinical questions are introduced to illustrate how machine learning might help solve them-perhaps bringing anesthesiology into an era of machine-assisted discovery.

Figure 1:

Examples of model fitting to data. The data are synthetic, for the purposes of illustration only. (A) An under-fit representation of the data. Although the linear discriminator captures most of the green circles, numerous red crosses are misclassified. The linear model is too simple. (B) The discriminator is over-fit to the data. Although there are no classification errors for the example data, the model will not generalize well when applied to new data that arrives. (C) A parabola discriminates the data appropriately with only a few errors. This is the best parsimonious classification.

Figure 2:

Examples of model fitting using augmented variables. The data are synthetic, for the purposes of illustration only. (A) An example of a model-fitting problem in which desirable outcomes (represented by green circles) are clustered around a mean point, and adverse outcomes (represented by red crosses) are associated with deviations from that point. For clinical correlation, one might imagine that the data represent favorable or unfavorable ICU outcomes based on rigorous control of potassium (Feature K) and glucose (Feature G). (B) Rather than attempting to fit outcomes solely to the variables K and G, the variable space can be augmented by also fitting to K² and G². This example demonstrates that the fitting of a perimeter around a mean value is easily accomplished by a linear fitting within the augmented space of K² and G². The linear discriminant of (K²) + (G²) – 9 = 0 as shown produces a circular boundary of radius 3 in the K,G space.

Figure 3:

(A) The simplest, fully-connected neural network from two input features to one output. The weights for each connection are illustrated, and each neuron in the network uses the sigmoid activation function to relate the sum of its weighted inputs, z, to its output. The sigmoid function is $\sigma \left(z\right)={\left(1+e^{–z}\right)}^{–1}$ (B) Other biologically-inspired activation functions are possible and have practical benefits beyond the original sigmoid. Further evolutions are the tanh function (essentially two sigmoids arranged symmetrically), the softplus (the integral of the sigmoid), and the rectified linear unit (a non-smooth variant of the softplus).

Figure 4:

The property of universality means that neural networks can represent any continuous function. The neural network shown here represents a hypothetical system to take a photographic image of a patient and render a prediction of their Cormack-Lehane view at intubation. (Not all nodes and connections are illustrated, as the input and hidden layers would each contain several thousand nodes. More pragmatic network topologies can be applied to visual recognition problems than the general case shown here.)

Figure 5:

Recurrent Neural Networks (RNNs) employ feedback such that the output of the system is dependent on both the current input state and also the preceding inputs, enabling the network to respond to trends that evolve over time. In the Elman network arrangement shown here, the Context Layer feeds from and to the Hidden Layer.