Bayesian instrument for spinal disability: A predictive model for aircrew flying fitness

INTRODUCTION

Spinal disabilities occur in aircrew at the same rate as the normal population, with or without obvious pathology in the spine. Precious man hours are however lost in the diagnosis, treatment, rehabilitation, and assessment for fitness of aircrew each year. Due to the specialized nature of the job, the criteria for assessment and subsequent fitness to fly in the case of aircrew are different from the general population.^[1]

Department of Human Engineering, Institute of Aerospace Medicine (IAM) evaluates IAF aircrew with spinal disabilities before they can fly again. Clinical, radiological, and functional evaluations are performed, including the evaluation of aviation-specific simulators. These assessments have proven effective in reflighting aircrew or identifying those who need sheltered duties. This evaluation method has evolved over time. Empirically, it seems to classify aircrew who should return to flying and those who need more management or recovery time.^[1-3] Cross-sectional data from the Department of HE show that at any given point in time, while 57% of the aircrew with spinal disabilities may be awarded a flying category, 43% are sent back with a non-flying category.

It is important to standardize this procedure, identify the relative importance of each step, and eliminate redundant evaluation stages. This would improve the ability to predict the outcome of aircrew with a spinal disability and expedite their return to flying, maximizing trained manpower.

This study aimed to develop a Bayesian instrument for spinal disability (BIS) based on clinical, radiological, and functional evaluations of IAM aircrew. The composite model indicates the relative importance of ongoing evaluations and predicts focus areas affecting aircrew’s return to flying. It is also a predictive tool that may be used in the field before referral to IAM.

MATERIAL AND METHODS

One hundred twenty aircrews reporting spinal-related disabilities to the Department of Human Engineering, IAM participated in the study. Aircrew with disability in addition to that of the spine was not included in the study as these disabilities are independent predictors for return to flying status.

The aircrew documents available before evaluation at IAM, were perused to tabulate the following data:

1. Age

2. Gender

3. Branch

4. Stream

5. Disability

6. History of low back ache at the onset of disability

7. Mode of onset of disability

8. Salient magnetic resonance imaging (MRI) findings on the latest MRI

9. Modality of treatment (whether surgical or conservative)

10. Duration of the disease (days since onset of the disease)

11. Present medical category in which reported to IAM (whether presently in flying or non-flying category).

Following this, the opinion of clinical specialist was sought. Thereafter, their perceived functional status was recorded on short form 36 health survey also known as SF-36 form.^[4,5] The SF-36 is an approved measure of the medical outcomes trust, a highly regarded clearinghouse for outcome measures with proven psychometric properties and utility. The SF-36 is subdivided into 2 separate health constructs: The physical component summary score and the mental component summary score. Subscales of the SF-36 measure 8 different health concepts including general health, physical functioning, role functioning, bodily pain, mental health, emotional functioning, vitality, and social functioning.

They underwent functional evaluation as per protocol and opinion of the Aerospace Medicine specialist was then rendered as a binary outcome variable of either flying or ground medical category.

The collected predictor variable data set was used to construct Bayesian Networks (BNs) for classification of the aircrew into an outcome of final medical category of flying or grounded. The BNs were created using GeNie Modeler, developed at the Decision Systems Laboratory, University of Pittsburgh.^[6,7]

RESULTS

Demographic data

1. 83.3% of the aircrew were below 40 years of age.

2. Only seven aircrews were female out of the total 120 aircrews in this study.

3. 89.2% of the aircrew were flying pilots. 46.7% of the aircrew in this study were from the helicopter stream.

4. 55% of cases of spinal disability had a documented prolapsed inter-vertebral disc (PIVD). However, more significantly, 40.8% had no clear evidence of radiological abnormality or trauma as the cause for the back pain. 67% of the aircrew in this study had an insidious onset of back pain without being related to a specific traumatic or systemic cause.

5. The latest MRI of the aircrew showed no nerve root compression in 79.2% of the cases.

DISCUSSION

All the predictive factors were analyzed and the relationships between factors were modeled using the Augmented Naïve Bayes algorithm to decide the final fitness of the aircrew.^[9-11] The acyclic graph of the BIS with the current beliefs based on the data of 120 aircrews is presented in Figure 1.

Close

Figure 1:

Acyclic graph showing the current beliefs of all nodes. Note: In this graph, all the variables or “nodes” are depicted with their interrelationships to each other. The inter-relationships or influences are depicted using arrows or “arcs.” The arcs are directional in nature. An arrow leading from a node to another node means that the first node is influenced by the state of the second one. The strength of influence of one node on the other is a function of the conditional probability tables.

Export to PPT

Validation of the BIS was carried out using the LOO method. In this, the network was trained on 119 records (n-1) and tested on the remaining one record. The process was repeated 120 times. The model structure was kept fixed, but the conditional probabilities were learned using 119 data points. The resultant network was then tested using the data of the left-out data point. The accuracy was found to be 0.97, sensitivity was 0.97, and specificity was 0.96.

The BIS was also tested on partial data from 29 aircrew. Three important variables/factors, which are generally collected once the aircrew reports to IAM, were excluded from the database. The excluded factors were Clinician’s Opinion, SF 36 Questionnaire score, and the functional evaluation outcome. This data were used to explore the utility of the Bayesian model based on data that are available before the aircrew reports to IAM. The results showed that:

1. The results show that out of 29 cases, the BN was able to correctly predict the outcome variable state (Flying Category or Ground Medical Category) 25 times. This means that the model has an accuracy ([True Positive + True Negative]/[Total Population]) of 0.86.

2. It has a sensitivity ([True Positive]/[Actual Positive]) of 0.94 wherein it correctly predicted 16 cases as flying fit out of the 17 cases who had been deemed flying fit as per the data.

3. It has a specificity of (True Negative/Actual Negative) of 0.75 wherein it correctly predicted 9 cases as ground medical category out of 12 cases who were recommended ground medical category as per the data.

Implications of the BIS

1. Software version of the BIS. The predictive model/scale has been developed on an open-source platform GeNie Modeler, developed at the Decision Systems Laboratory, University of Pittsburgh, and licensed to BayesFusion, LLC. Academic license for use for academic purposes is available with the authors. Since IAM is the only center for the evaluation of spinal disabilities in the IAF, the BIS is being used at the Department of Human Engineering at IAM.

2. Individual Scoring and Prediction. The BIS uses demographic characteristics such as gender, age, and stream in its predictive model. It also takes into consideration the specific diagnosis, radiological findings, and treatment. Thus, the BIS scores an individual and gives his probability of upgradation to a flying category based on factors unique to him. This enables follow-up of improvement on an individual basis. It also enables data-based differentiation of final outcomes between two individuals who may be similar in certain aspects of their disability only.

Case study on an individual aircrew. The spinal disability scale applied in one case shows a 95% probability of getting a flying medical category [Figure 2]. Figure 2 shows that in the given case, there is a definite value or state associated with each variable which is a predictor in the spinal disability scale. This particular case is of a 30–40-year-old male aircrew who had reported within the past 1 year with low back ache which was not associated with any trauma or bony lesion. He has no neurological findings on his MRI and has been recommended a flying category by the neurosurgeon. He has scored fairly high on the SF 36 indicating adequate functional status before reporting to IAM. His Human Engineering evaluation is normal. In this case, there is a 95% probability that this aircrew would be considered fit for flying.

Close

Figure 2:

Outcome prediction of a single case using the Bayesian instrument for spinal-disability model Note: The figure shows that in the given case, there is a definite value or state associated with each variable which is a predictor in the spinal disability scale. In this case, there is a 95% probability that this aircrew would be considered fit for flying.

Export to PPT

CONCLUSION

BIS was constructed using a BN from the data collected of the 120 aircrew who reported to IAM for evaluation. It considered the clinical, functional, and radiological outcomes to predict an aircrew’s fitness during the evaluation at IAM. It takes into consideration various demographic factors including the stream of the aircrew. The spinal disability scale gives individual scores/probabilities of the outcome and can direct targeted follow-up. The BIS showed high accuracy, sensitivity, and specificity during validation. A software-based implementation has been developed and placed in the Department of Human Engineering at IAM.