2023-09-12 03:24:14
Why is this important?
Bariatric surgery is a treatment that can be offered as a second step to patients who are morbidly obese or associated with high-risk comorbidities, in order to promote weight loss, limit complications and improve life expectancy. However, following the surgical procedure, the result is very heterogeneous from one patient to another, and difficult to predict individually. However, the existence of a predictive tool would help practitioners and their patients in evaluating the benefit-risk ratio of surgery. Analytical models have been developed previously, but do not predict long-term weight loss, or are not statistically powerful enough. This work, carried out by French teams, is one of the few to have used AI to predict post-surgical weight loss and to have validated its performance using several cohorts and randomized trials.
Methodology
The predictive model was developed from two prospective longitudinal cohort studies following adult patients following a first bariatric surgery (ABOS and BAREVAL), using a convolutional neural network. It was then validated using eight cohorts (French, Dutch, Swedish, Italian, Indonesian, Brazilian and Mexican) and two randomized multicenter trials (SleevePass and SM-BOSS). All three types of approaches were included in the model: sleeve gastrectomy, adjustable gastric band and bypass gastric.
Principle results
In total, 10,231 patients (18-74 years old, 75.3% women) were included, or 30,602 patient years. At inclusion, the BMI was between 26.7 and 94.1 kg/m² and 28.2% had type 2 diabetes. The approach most often considered was bypass gastric, then the sleeve gastrectomy and gastric banding (65.4%, 28.1% and 6.5% respectively).
At 5 years, the median weight loss was 26.8%, with an overall identical trajectory of maximum weight loss reached between 1 and 2 years, followed by limited weight regain. It was significantly higher following bypass that following sleeve gastrectomy or following insertion of a ring (28.2% versus 23.6% and 14.9% respectively, p<0.0001 for both).
Among the 434 variables available, seven appeared relevant for predicting weight loss at 5 years in the two cohorts used for the development of the algorithm: preoperative weight, height, age, type of intervention, smoking history, diagnosis and duration of type 2 diabetes.
The predictive performance of the model was verified at 1, 2 and 5 years using the validation cohorts. At 5 years, the mean difference between predicted and observed BMI was -0.3 kg/m². Predictive accuracy was good at intermediate follow-up times.
Principales limitations
The panel of variables available for the development of the algorithm included little socio-economic, ethnic origin, behavioral or nutritional data and no genetic analysis data, even though they have an influence on the individual trajectory of people.
1694490842
#Artificial #intelligence #predicts #expected #weight #loss #bariatric #surgery
