ISSN 1470-3947 (print) | ISSN 1479-6848 (online)

Endocrine Abstracts (2019) 63 P259 | DOI: 10.1530/endoabs.63.P259

Neural network for predicting recurrence of the Cushing's disease within three years after neurosurgical treatment

Elena Nadezhdina, Olga Rebrova & Andrey Grigoriev


Endocrinology Research Center, Moscow, Russian Federation.


Introduction: The reccurence rate after successful neurosurgery for patients with Cushing’s disease (CD) varies between 10% and 65%. It is shown earlier that some features are associated with the probability of recurrence, yet no rules were elaborated how to analyse the set of variables.

Materials and methods: The retrospective study was based on 219 cases of CD (32 men, 187 women) with a disease duration ranged from 4 months to 22 years, who underwent an endoscopic transsphenoidal adenomectomy in 2007-2014 at the age of 38±12. The duration of a postoperative follow-up was 3 years or more. The 3-years remission persisted in 172 patients, the recurrence developed in 47 patients. Univariate and multivariate analyses were performed.

Results: The univariate analysis did not reveal association between the CD recurrence and sex, age, duration of disease, size of adenoma. At the same time, adrenocorticotropic hormone (ACTH) and cortisol levels in the early postoperative period found out to be statistically significant predictors: patients with ACTH levels less than 7 pg/ml had recurrence about 6 times less than patients with higher levels of ACTH, odds ratio (OR) 0.16, 95% CI [0.06; 0.43]; patients with cortisol levels less than 123 nm/l had recurrence about 5 times less than patients with higher levels of this hormone, OR 0.19 [0.09; 0.39]. However, in 58% of cases the levels of these two hormones are dissociated, so it is not possible to use any of them to predict recurrence. This made multivariate analysis reasonable. No effective logit-regression model was developed, but the highly effective neural network (3-layer perceptron) was obtained. The predictors are sex, age, duration of disease, size of the adenoma, levels of ACTH and cortisol in the blood in the early postoperative period. The correct prediction on randomly generated samples was 94.2% on the training sample (n=155), 84.4%, on the control sample (n=32), and 87.5% on the test sample (n=32). The model has sensitivity 74%, specificity 97%, positive predictive value 85%, negative predictive value 93%.

Conclusion: The developed ANN is an effective tool for predicting recurrence and it will allow to perform personalized approach to the management of patients who underwent neurosurgery, eg. vary frequency of dynamic examinations, prescribing and correcting drug therapy, etc.

Article tools

My recent searches

No recent searches.