Real-time prediction of an anesthetic monitor index using machine learning Olivier Caelen a , Olivier Cailloux b , Djamal Ghoundiwal b , Abhilash Alexander Miranda a , Luc Barvais b , Gianluca Bontempi a a Machine Learning Group, D´ epartement d’Informatique, Facult´ e des Sciences, Universit´ e Libre de Bruxelles, Bruxelles, Belgium http://dev.ulb.ac.be/mlg/ b Service d’Anesth´ esiologie-R´ eanimation, Facult´ e de M´ edecine, Universit´ e Libre de Bruxelles, Bruxelles, Belgium Abstract An anesthesiologist may control the level of consciousness of a patient undergoing surgery by appro- priately dosing hypnotic drugs. The information provided by the monitoring devices may be utilized in order to accomplish this task. One such monitor provides a dimensionless quantity derived from the electroencephalogram called bispectral index (BIS), which could quantify the level of awareness of the patient. This article discusses the use of machine learning techniques to implement a predictive model of the BIS based on the variation of the hypnotic drugs. Such a model learned from a database of recorded operations can aid real-time decision making during the course of an operation. In order to deal with inter-individual variability, the proposed model takes into account patient physiology as well as the reactions of the patient during the early phases of the operation. Two models of the bis- pectral index behavior are assessed and compared in this work: a linear predictor and a local learning predictor. These prediction models were software implemented and their accuracies were assessed by a computerized cross-validation study and were tested in real situations. Key words: anesthesia, bispectral index, machine learning, local modeling 1. Introduction During surgery, the anesthesiologist controls the depth of anesthesia by administrating three types of drugs: hypnotics to cause and maintain loss of consciousness, analgesics to inhibit pain, and very often muscle relaxants to block mus- cle reactions. In this paper, the drugs consid- ered are propofol as hypnotic and remifentanil as analgesic. Nowadays, anesthesiologists may take advantage of devices which monitor unconscious- ness in real-time in order to choose the appropri- ate dose of hypnotics. Typically, such monitors are connected via electrodes to the patient’s fore- head and display a signal that has been derived from the electro-encephalographic activity of the patient. The value of the signal gives the anesthe- siologist an indication of the level of unconscious- ness of the patient. A commonly used monitor is the bispectral index (BIS) commercialized by Aspect Medical Systems [39]. The BIS monitor provides a single dimensionless number, the BIS value, which ranges from 0 to 100. A BIS value of 0 equals EEG silence, while near 100 is the ex- pected value for a fully awake adult, and between 40 and 60 indicates a level for general anesthe- sia recommended by the manufacturer. Figure 1 shows a typical temporal pattern of the BIS sig- nal during a surgical operation. The BIS signal is close to 100 at the beginning of the operation when the patient is still conscious and falls to about 50 after the induction phase when the pa- Preprint submitted to Artificial Intelligence in Medicine June 7, 2011