Our study consisted in a retrospective single center study of IONM and epidemiological data of patients undergoing neurosurgery during the years 2020–2022. The local Institutional Review Board exempted this prospective analysis of retrospectively collected data as research.
The inclusion criteria were patients admitted for cranial surgery with neuro-oncology (primary and secondary tumours) and/or neurovascular pathology (arteriovascular malformations, aneurysm clipping, cavernoma and arteriovenous fistulas). Both adult and pediatric patients were included. Patients were grouped in 4 groups according to their age, as seen in Table 1.
Table 1
Demographics of the included patients
AGE
|
SEX
|
LOCATION
|
Group 1: 11 patients
|
Female: 6
Male: 5
|
Frontal: 3
Parietal: 6
Temporal: 2
|
Group 2: 31 patients
|
Female: 12
Male: 19
|
Frontal: 14
Parietal: 11
Temporal: 5
Middle Cerebral Artery (MCA): 1
|
Group 3: 72 patients
|
Female: 37
Male: 35
|
Frontal: 29
Parietal: 12
Temporal: 14
Insular: 2
Anterior Communicating artery (ACom): 2
Posterior Communicating Artery (PCom): 2
MCA: 11
|
Group 4: 129
patients
|
Female: 68
Male: 61
|
Frontal: 33
Parietal: 37
Temporal: 25
Insular: 5
Occipital: 8
ACom: 5
MCA: 15
Interal Carotid Artery (ICA): 1
|
Group 5: 24 patients
|
Female: 14
Male: 10
|
Frontal: 3
Parietal: 10
Temporal: 4
Occipital: 2
MCA: 3
Acom: 2
|
Demographics for age, sex, and location of the lesions as well as preoperative and postoperative neurological function were collected from electronic records and are shown in Table 1. IONM techniques were divided into two categories: mapping and monitoring. For both procedures, two types of electrodes can be used: direct cortical/subcortical electrodes and transcranial electrodes. All signals were recorded using built-in amplifiers and filter tailoring was performed to record specific signals by a Medtronic Nim Eclipse system® with IONM software.
Assumptions based on the existing literature were made before building the Bayesian Network. We assumed that transcranial stimulation and recordings are less specific than direct cortical in brain surgery. This is because the distance from the electrode to the functional brain area at risk is larger when recording or stimulating through the skull (11). We also assumed that multimodality IONM should be preferred to unimodal (12). Based on this, we classified IONM procedures into 3 groups:
1. Optimal intraoperative neuromonitoring:
Multimodality IONM including direct cortical and subcortical Motor Evoked Potentials (MEPs), direct cortical Somatosensory Evoked Potentials (SSEPs), direct cortical language mapping, direct cortical Visual Evoked Potentials (VEPs), electrocorticography (EcoG) and all their transcranial modalities when possible. Monitoring is done continuously to guarantee real-time signal acquisition during surgery to assess functional integrity of the brain.
Suboptimal intraoperative neuromonitoring.
Multimodal IONM but exclusively transcranial modalities or cortical modalities performed and not in a continuous manner.
Inadequate intraoperative neuromonitoring.
Unimodal IONM or IONM that did not guarantee safety for the patient. For example, sporadic signal acquisition.
Significant Signal Changes
For MEPs, the following signal changes were considered significant: (1) an abrupt disappearance when stimulating at threshold intensity, defined as an intensity that elicited at least 50% of the MEPs from the muscles after repeated stimulation of the motor cortex of the brain; (2) changes in stimulation threshold eliciting MEPs (above 100V if transcranial, above 3mA if direct cortical); (3) a reduction of more than 80% of baseline signal amplitudes at supramaximal intensity, defined as the intensity able to elicit 100% of the MEPs at baseline after repeated stimulation (13); and (4) an approximate mapped distance of less than 5mm to the corticospinal motor tracts during tumor resection surgery, estimated by dynamic subcortical MEPs (14). For SSEPs and VEPs, we considered a reduction of 50% amplitude and 10% increase in latency of baseline signals (15,16). Finally, for speech we considered difficulties encountered during surgery, for example, difficulties in initiation of speech or fluency (17).
We further classified signal changes as:
1. Reversible, if the signal change reverts after surgical corrective action is taken when brain damage is suspected in surgery. In this case, recovery of the functional area is expected. Corrective actions we considered included taking a surgical break or stopping resection in that area, blood pressure increase, warm irrigation or papaverine administration in some cases.
2. Irreversible, if the signal change from baseline does not revert, regardless of whether corrective actions were taken. An irreversible signal change has high chance of permanent damage to the neural structure related to this signal.
Bayesian Network Approach
Bayesian Networks are directed acyclic graphs. A graph is a mathematical abstraction that consists of nodes and edges that connect the nodes. As acyclic graphs, Bayesian Networks do not permit cycles between its nodes. There are 3 different types of nodes in Bayesian Networks: chance nodes, decision nodes and utility nodes. In addition, Bayesian Networks contain a set of conditional probability distributions that represent the relationship of its nodes and their potential causal dependence, represented by the network´s edges. The edges are based on expert domain knowledge and are often assumed based on logistic regression associations between the nodes. The edges in a Bayesian Networks define every possible outcome of the preceding causal nodes in the form of conditional probability distributions. Given observed evidence, the prior probability of that evidence is defined as the probability of its occurrence before new data is collected. Bayesian networks use prior probabilities to estimate the conditional probability distributions of each of its nodes, in the presence of new, unobserved data.
We used the commercial software AgenaRisk® to implement our network. We incorporated expert domain knowledge from our own working experience and the literature to design the network and establish its nodes and edges (Fig. 1). The network´s prior probabilities were computed with data from retrospective cases performed in our center, for which we knew the outcome. With this data, we first analyzed the original state of our prior probabilities. The original state included the prior probabilities of patients having optimal IONM, the probability that a neurophysiologist identifies a signal change and whether this was reversible or permanent change, the probability of surgeons causing significant damage to the patients overall and the probability that patients with a signal change wakes up with a new deficit. Secondly, we looked at counterfactual reasoning. Counterfactual reasoning was calculated by filling potential new observations in the network’s nodes, over the computed prior probabilities.
Our aim was to answer the following questions:
1. What is the likelihood of patients waking up intact, with a transitory deficit or with a permanent deficit if a reversible signal change was seen in surgery and surgeons took corrective action vs when they did not?
2.What is the outcome if a permanent signal change is observed?
3. If a patient wakes up with a permanent/transitory deficit or intact, what is the likelihood that a permanent/reversible signal change, or no change was detected? What is the likelihood that corrective action was taken?
4. What is the decision that maximizes the utility when a signal change is detected (both transitory or permanent), taking corrective action or not taking it?
Software packages such as AgenaRisk® can update the probabilities in a network using new observations based on the network’s architecture and through the application of Bayes’ theorem
$$P\left[\text{C}\text{a}\text{u}\text{s}\text{e}|\text{E}\text{v}\text{i}\text{d}\text{e}\text{n}\text{c}\text{e}\right]=P\left[Evidence|\text{C}\text{a}\text{u}\text{s}\text{e}\right]*\frac{P\left[Cause\right]}{P\left[Evidence\right]}$$
1
where P[Cause] is the prior probability of a cause, P[Evidence] is the prior probability of observing a piece of evidence, P[Cause|Evidence] is the probability of a cause given an observed evidence and P[Evidence|Cause] is the probability of observing a piece of evidence in the presence of the cause. Our network architecture had the nodes and edges that are described subsequently.
Chance Nodes
Table 2 shows the chance nodes we considered in our network. i.e. the nodes whose outcomes don´t depend on the decision maker.
Table 2
Chance nodes in the network.
Chance Nodes
|
Description
|
Age (years)
|
Group 1: 0–18
Group 2: 19–30
Group 3: 31–50
Group 4: 51–70
Group 5: 71–100
|
Adult/Pediatric
|
Adults were patients over 18 years.
|
Sex
|
Male or Female
|
Type of Lesion
|
Cortical (Brain tumor) or Vascular (Aneurisms, Arteriovenous Malformations)
|
Location
|
Tumors and AVMs: Frontal, Parietal, Temporal, Occipital and Insular
Aneurisms: Middle Cerebral Artery, Internal Carotid Artery, Anterior Communicating Artery (ACom) and Posterior Communicating Artery (PCom).
|
Preoperative motor deficit
|
No, Minor weakness (3–4/5 in the MRC muscle power scale), Major weakness (0–3/5).
|
Preoperative Sensory Deficit
|
No, Minor deficit (paresthesia), Major deficit (anesthesia).
|
IONM Discrepancies (Between Neurophysiologist and Neurosurgeon)
|
None: Neurophysiologist offered modalities match modalities used by neurosurgeon.
Minor: One or two modalities offered are missing.
Important: Important modalities offered are missing.
Crucial: Surgeon declines most modalities offered.
|
Modalities
|
Optimal, Suboptimal, or Inadequate.
|
Anesthesia
|
Total Intravenous Anesthesia, Halogenated Agents (Sevoflurane, Isoflurane), Others (Dexmedetomidine, Nitrous Oxide)
|
Baselines
|
Perfect, Imperfect but correlated to preoperative clinical status or Imperfect but uncorrelated.
|
Signal Change
|
No, Reversible, Irreversible
|
Postoperative Deficit
|
No, Transitory after neurorehabilitation, Permanent.
|
Decision Nodes
Only one decision node was considered, namely acting after a signal change. Corrective measures in response to a signal change include avoiding the surgical site or taking a surgical corrective maneuver after the warning (e.g. removing a temporary clip) taking a surgical break, raising the blood pressure to increase perfusion to the area, irrigation with warm saline, papaverine infusion in the suspicion of a vascular event. Based on this, we defined three types of corrective actions:
-
Adequate: all measures attempted.
-
Insufficient: some but not all measures were or could be attempted.
-
Inadequate: nothing was or could be done.
Utility Nodes
To define a suitable utility in IONM we identified two main options. The first option associates a subjective value to different outcomes such as complete surgical tumor removal despite a neurological injury, or preservation of neural function despite incomplete tumor removal. The second is to simply associate a monetary utility to the outcome. The latter option is easier to formulate, although it provides an incomplete assessment of the value of an outcome.
We quantified the potential net monetary utility of the presence or absence of a postoperative neurological deficit based on the healthcare costs incurred, which were estimated following the figures previously reported in [18, 19]. Patients with a postoperative deficit in our center are normally admitted to the neurorehabilitation unit to recover. On average, patient stay in neurorehabilitation spans from 8–16 weeks. We assumed that patients with a new onset deficit that recovered after treatment in the neurorehabilitation unit would need around 8 weeks. This would have a total cost for the National Health Service (NHS) of £56,952 for pediatric patients and £29,680 for adults. In contrast, if the deficit was permanent, these patients would have at least stayed 16 weeks in neurorehabilitation. This would cost the NHS £113,904 for pediatric patients and £59,360 for adults. In contrast to this, if a patient did not wake up with a deficit, the cost to the NHS would be £0. Also, if the deficit was present but recovered, the NHS would have saved the additional 8 weeks’ cost of admission, which would have been spent if the deficit was permanent. Costs, benefits, and net utility nodes were added to the network to account for these figures. The net utility was calculated as:
$$Net Utility=Benefits-Costs$$
2
where benefits are the total savings of the NHS (16 weeks´ savings if no deficit was present) and costs are the total costs (16 weeks´ admission in neurorehabilitation for a permanent deficit, 8 for a temporary deficit)
Expert domain knowledge prior probabilities
We used epidemiological prior probabilities published in the literature to build the nodes corresponding to preoperative motor and sensory deficit. We did not use epidemiological data for other nodes and chose to use data from our center instead. Regarding the nodes of preoperative motor weakness and preoperative sensory weakness for neuro-oncology, Motomura et al. studied the incidence of motor deficits in pediatric frontal tumors [20]. This was approximately 36%, of which 20% were mild. In adults, a similar study was carried out by Amidei et al. [21]. Motor deficits were present in 26% of the population. Regarding sensory deficits, Chandana et al. [22], reported that in adults these consisted of around 13% in parietal lobe tumors. No data could be found in the pediatric population. Regarding visual deficits, for tumors in the occipital lobe of the brain, Goodwin et al. [23], reported an incidence 11% of cases in adults and Peragallo et al. [24] a 27% in children. A total of 11% of the visual deficits were severe, for instance homonymous hemianopia. No data was found for motor deficits in occipital and parietal tumors. However, this area of the brain is not the primary motor area, so motor deficits are not expected.
A similar approach was taken regarding vascular lesions. Raps et al. [25] reported on the clinical spectrum of unruptured intracranial aneurism in adults. They found 11% of motor deficits and a 10% of sensory deficits. Regarding the pediatric population however, no data could be found on sensory deficits of unruptured aneurisms while a 53% motor weakness was reported [26]. Regarding visual deficits, the data mainly focused on posterior communicating aneurisms and was of 25% in the pediatric population and 29% in adults [27].
Expert Domain Knowledge Edges:
The causal inference edges proposed in the model are proposed in Table 3. They are the result of expert domain knowledge based on published association studies through frequentist statistical analysis.
Table 3
Edge
|
Evidence
|
Age and Sex to preoperative motor and sensory function as well as to type of lesion
|
Roper et al. 2005
|
Location of the brain area to preoperative motor and sensory function
|
Amidei et al. 2015, Katrib et al. 2018
|
Surgeon knowledge in IOM to modalities requested
|
Tamaki et al. 2021
|
Modalities requested to Baseline responses
|
Guérit et al. 2016
|
Preoperative Sensory and motor function to baseline responses
|
Kobayashi et al. 2021
|
Anesthetic agents used to baseline responses
|
Lo et al. 2006, Tobias et al. 2008, Deiner et al. 2010, Mirallave-Pescador et al. 2021
|
Baseline responses to intraoperative signal changes detected.
|
No change detected if baseline not recorded.
|
Surgical Action performed in the surgery to potentially provoke a signal change.
|
Damage to functional areas cause signal changes.
|
Surgical Action taken after a warning to reversibility of the Signal change
|
Skinner et al. 2014
|
Signal change to postoperative deficit
|
Skinner et al. 2014
|
Action taken to Postoperative Deficit
|
Skinner et al. 2014
|
Postoperative Deficit and adult or pediatric to Costs or Benefits
|
Turner Stokes et al. 2007 and 2012
|
Costs and Benefits to Total Utility
|
Definition of total utility
|