<aside> ⛏️
Development of a machine learning algorithm to predict cardiac arrest during operations in the operating theatre. Immersion in the operating rooms of the Toulouse University Hospital to carry out a complete engineering study with a team of 15 students. Final realisation using the random forest algorithm.
</aside>
The MIG FORENSIC studies decision support in real-time surgery according to two components, explored in parallel to propose an overall vision, namely :
While data science, "big data", statistical learning (Machine learning and Artificial Intelligence (AI) are now an integral part of the engineering landscape, the approach of delegating all or part of the decision to an algorithm is still not widespread in surgery. This can be explained on the one hand by the limited amount of data available; on the other hand, by the difficulty to develop a "responsible AI" in the broad sense, i.e. to answer the ethical questions raised by algorithms used for critical decisions but with relative reliability. The case study, proposed by the paediatric surgery centre of the university hospital of Toulouse, addresses these two aspects in concert. It consists in both : to first develop classification algorithms to predict cardiac arrests during paediatric surgery; and second to study how these tools may be accepted in a professional medium.
To study these two objectives, the group has been split up into two parts :
One group will study actual patient monitoring data, recorded during several months in six operating rooms. This corpus of data has unique characteristics that make it difficult to analyse: limited number of cases, multiple number of cases, multiple pathologies, varied population (from babies to young adults). It is about taking control of this data through the entire processing chain, in order to develop the first classification algorithms capable of predicting heart failure during major cardiac insufficiencies likely to seriously disrupt the course of the surgery or even threaten the life of the patient. Courses signal processing, statistical learning and software engineering in the first week have allowed a progressive approach to this field.
The other group has conducted non-participatory observations in operating theatres (Toulouse University Hospital, paediatric surgery centre). The objective was to model the decision process during the surgical act, and to determine under which conditions and to what extent this decision could be assisted by a tool based on statistical learning. The experts' confidence in such a tool depends in particular on its long-term reliability, and its intelligibility when predictions contradict intuition. However, most of the time, the performance of classification algorithms increases at the cost of a significant complexity, until they turn into "black boxes". This is particularly true of Artificial Intelligence algorithms, which have achieved the performance of the best experts. The students thus had to answer the following question: how can we consider introducing a solution into operating rooms that would violate the experience of the doctor ?
In the hospital, very little data is collected, and most of the time it is impossible to analyze it properly due to the lack of unity between the different sources. During the past decade, efforts have been made to collect data at a higher scale and to centralize it, but there is still a very long way to go.
However, data analysis is ubiquitous in medical research. During a visit to the research center in cancerology of Toulouse, several research teams, including a biologist or a doctor and a data scientist presented us the way they use machine learning algorithms for their work. The quality of the data is good and allows them to carry high level research work. Nevertheless, the situation when it comes to treatment is far from this excellence.
In the operating room, data from the monitor is usually only available at the moment and is not saved. Instead, the evolution of the parameters is reported on an anesthesia sheet that is kept. Information technology can sometimes be overlooked in the medical field, thus it is relevant to wonder if data has a role to play in the operating room.
Our role in the project was manyfold, we have enlightened data for the data team, we have questioned the goal of the project in regard to the results we had, we defined the final tool in order to design it to implement it in operating rooms and we discussed the potential adverse consequences of the tool and issues its implementation raises.
How can the solution be implemented in the operating room ?
In order to be accepted, we asked anaesthetists about what a good tool is for them, and their answer was :
As anaesthetist are constantly using sound to be alerted, the solution must include an audible alarm.
In the surgical service of the pediatric hospital, there are 9 operation rooms. Usually 5 of 6 are used simultaneously. Each one has a speciality: cardiac surgery, orthopedics, visceral surgery, etc. The regulators make the schedule for all the doctors, interns and nurses.
Before the patient’s arrival, the anesthetic team (usually made up either by one intern or a nurse and an attending anesthesiologist who supervises the latter) sets up the tools for anesthesia. They agree on a care protocol for the patient, ~~~~and set up the monitoring devices.
The patient arrives and is transferred on the operating table. The anesthetic team sets up electrodes on their torso and a SpO2 sensor on one finger. Young children are put down to sleep by the anesthetic team using a special gas. Only then, an IV is set up and they are being injected Propofol to keep them asleep. Older children are directly being injected Propofol, and the IV is set up before sleeping. Analgesics are alwayes used, while paralyzing agents like curare are not always needed. When the patient is asleep, more invasive sensors can be set up, like in the femoral artery. The patient can also be intubated; it is usually the case for newborns.
When the anesthetic team is done with the patient, the surgeon is called. The information about the patient is checked. The surgeon incises the operating site and goes on with surgical gestures. They are helped by the intern and the a nurse. The three are scrubbed in while the rest of the people in the OR are not. During the entire surgery, the patient is monitored by a member of the anesthetic team – an intern or a nurse if the surgery is basic, the attending physician if the surgery is risky. When the surgery is basic, the attending physician navigates between circa three operating room. Monitoring devices give information like cardiac rhythm, pressure, body temperature, SpO2 and pCO2. The anesthetic team sets up alarm threshold for each of those parameters. When an alarm goes off, the anesthetic team silences it and analyzes the situation. Alarms are frequent and most of the time without gravity.
When the surgeon is done, they leave the operation room and the anesthetic team takes the lead. The anaesthetist and their intern will agree on a protocol to wake up the patient and post-surgery care. Critical patients are directly transferred to PICU or NICU. Otherwise, general anesthesy is interrupted, but analgesics are maintained. The anesthetic team looks after the patient during awaking, and stimulates them during the last stage. As soon as the patient shows signs of consciousness, they are conducted to the recovery room.
If another surgery is scheduled right after, the remaining of the anesthetic team prepares the operation room for the next patient. Special caregivers thoroughly clean the operating room while the scrub nurse removes all the sterilized elements previously used.
On the course of discussions with the practitioners, some requests emerged :
Attending operations allowed us to propose a model of the interactions inside the operating room. Usually 4 to 6 people are in the room: a surgeon, an anaesthetist nurse, other nurses, and sometimes a surgery student and an anaesthetist. Indeed, an anaesthetist deals with several operations at the same time thus he is not always present in the room. The room may be divided in two poles: the anaesthesia pole and the surgery pole. In the anaesthesia part, located next to the patient’s head, the anaesthetic nurse can watch the patient’s electrocardiogram, O2 saturation, blood pressure and temperature on a screen. The artificial respirator and the medicine that the patient might need are also located there.
Except for the anaesthetist announcing when the induction is finished so that the surgeon can start the operation, and the surgeon announcing the end of their intervention so that the anaesthetist can start waking up the patient, the two poles rarely communicate.
.jpeg)
At first glance, the prediction algorithm our team is developing is supposed to be used by the anaesthetist who is responsible for the patient’s general well-being during the time of the operation. They are responsible for the induction and awaking phase. They control the anaesthesia with the artificial respirator, and medicine delivered through perfusions. They spend time watching the patient’s ECG, oxygen saturation, arterial pressure, capnia, respiratory frequency, but also the patient’s appearance, especially the way they breath and the color of their face.
.jpeg)
Our observation showed that the anaesthetists are the ones who receive information about the patient and know how to deal with contradictory information and how to react. The algorithm should then have the same place in the room as the data from the monitor, that is to say alerting the anaesthetist by collecting data from the patient.
However, the decisional process of the anaesthetist is supposed to be modified by the system. We imagined three possible reactions from the anaesthetist: indifference, anticipation and action.
.jpeg)
Question of the reaction of the medical team
One thing that is important is the reaction of doctors, especially for the purpose of the implementation of a device that challenges the current paradigms of medicine. Indeed, our tool may seem complicated or even obscure to some health workers. Thus, we need to ensure that there is trust between the medical team and the output of the algorithm.