Artificial intelligence approaches may help identify acromegaly earlier and support more individualized prediction of outcomes after surgery, medical therapy, and radiotherapy, although broader validation is still needed before routine clinical implementation, according to a review published in The Journal of Clinical Endocrinology & Metabolism.

The authors reviewed applications of artificial intelligence (AI) across the acromegaly care pathway, focusing on disease detection and treatment management. They searched PubMed for English-language articles using the terms “artificial intelligence AND acromegaly,” and also considered selected papers cited by identified articles. 

Facial recognition. The review highlighted AI-assisted analysis of facial photographs as one potential approach to addressing the 6- to 10-year diagnostic delay commonly reported in acromegaly. These systems use machine learning (ML) and deep learning to identify facial characteristics associated with the disease, including geometric and visual or textural features. 

In one early study from China, an ensemble model was developed using facial photographs from 527 patients with acromegaly and 596 controls. It achieved 96% sensitivity, 96% specificity, a 96% positive predictive value, and a 95% negative predictive value, outperforming specialist and primary care physicians who evaluated the same images. 

Subsequent studies evaluated deep learning and transfer learning. One model analyzing photographs from 1,131 patients achieved 95% diagnostic accuracy. A multicenter Swedish analysis provided a comparison with physicians. Six ML models were trained using smartphone facial images from 155 patients with acromegaly, 79% of whom were biochemically controlled, and 153 matched controls. The best-performing facial recognition model had 82% sensitivity and 87% specificity, compared with 66% sensitivity and 93% specificity among 12 experienced endocrinologists. 

Another study examined whether AI could recognize acromegaly before clinical diagnosis. Investigators analyzed 489 photographs from 92 patients with acromegaly and 254 images from 88 controls, including photographs taken 3, 6, 9, and 12 years before diagnosis. The model performed best on prediagnosis photographs, taken an average of 7.47 years before diagnosis, achieving 92% accuracy.  

Hand images. AI-based detection has also extended beyond facial photographs. A multicenter Japanese study used 11,480 hand images from 317 patients with acromegaly and 399 controls across 15 pituitary centers. A deep learning model achieved 89% sensitivity, 91% specificity, 88% positive predictive value, and a 93% negative predictive value, with an F1 score of 0.89 compared with a range of 0.43 to 0.63 among specialists. 

Voice analysis. Digital voice analysis also differentiated patients with acromegaly from controls. In a Swedish study of 151 patients with acromegaly and 139 matched controls, ML models analyzed 3,274 acoustic parameters derived from a sustained vowel. The model achieved an area under the receiver operating characteristic curve (AUROC) of 0.84 compared with 0.69 for experienced endocrinologists. 

Surgical remission. Beyond detection, the review described ML models designed to predict treatment outcomes. For surgery, a model developed in 668 patients used eight variables to predict biochemical remission after transsphenoidal surgery, achieving an AUROC of 0.86 in the training cohort and 0.82 in the validation cohort. The model underwent internal but not prospective or external validation. Another surgical study developed a preoperative model and a full model incorporating preoperative characteristics, surgical findings, and postsurgical hormone levels. The models were developed in 833 patients and subsequently evaluated prospectively in 99 patients and externally in 52 patients. The full model achieved an AUROC of 0.87 in both the development and prospective validation cohorts. Postoperative day 1 growth hormone level, extent of resection, and Knosp grade were its most influential predictors. 

Medical treatment. ML also showed potential for predicting response to first-generation somatostatin receptor ligands (fg-SRLs). In a multicenter cohort of 153 patients, six algorithms were evaluated using clinical, biochemical, and histopathological variables. The best-performing support vector machine incorporated age, sex, pretreatment growth hormone and insulin-like growth factor I levels, somatostatin receptor subtype 2 and 5 expression, and granulation pattern. It achieved 86% accuracy for predicting biochemical control during fg-SRL therapy, compared with 55% for somatostatin receptor subtype 2 expression alone. 

Radiotherapy. Evidence for predicting radiotherapy outcomes was more limited. In a study of 57 patients, a radiomics signature achieved an AUROC of 0.92 for distinguishing remission from nonremission. Combining radiomics with clinical parameters increased the AUROC to 0.96, compared with 0.86 for the clinical model alone. A separate deep learning model predicting remission after stereotactic radiosurgery achieved 93% accuracy in its training cohort and 85% in an external cohort. 

The authors identified several limitations to AI-based detection. Most facial-recognition studies included fewer than 200 patients and drew participants from a single or small number of tertiary referral centers, increasing the potential for overfitting and limiting assessment across subgroups. Most cohorts consisted predominantly of Asian and White populations, and differences in facial anthropometry across ancestries could affect model performance. Use of unmatched or noncontemporaneous controls could introduce differences in image quality, lighting, and demographic composition, while camera angle, resolution, makeup, and facial hair could also affect performance. Privacy and regulatory requirements present additional barriers to facial and voice recognition systems. 

“Early diagnosis and fast disease control are important to improve the care of patients with acromegaly," wrote the authors. They concluded that AI is promising for both goals and has the potential to serve as an adjunct to existing clinical tools for earlier detection and for moving treatment away from a trial-and-error approach toward precision medicine. 

Leandro Kasuki, MD, PhD, reported receiving speaker fees from Recordati and Ipsen. The other authors reported nothing to disclose.