Obesity treatment could become more precise with AI risk tools

Globally, about one in eight people is living with obesity, but health systems remain ill-equipped to deliver continuous, specialist-led care at the scale the disease now demands. New evidence suggests that AI could help doctors identify high-risk patients earlier, tailor treatment beyond body mass index, and extend behavioural support outside clinic walls. However, the same evidence also points to a major warning: AI tools for obesity management are advancing faster than the systems needed to validate them, audit them for bias, and monitor their safety after deployment.

The study, titled "Artificial Intelligence-Based Risk Stratification in Obesity Care: From Diagnosis to Personalised Treatment Pathways" and published in the journal Diagnostics, sheds light on recent AI applications in obesity care from January 2024 to January 2026 and assesses how ready these tools are for real-world clinical use.

Obesity care faces a scale problem that AI may help address

According to the research, obesity is a chronic, relapsing disease that cannot be managed effectively through brief, occasional clinic visits alone. Standard care often depends on episodic appointments, limited consultation time and fragmented services. Yet patients frequently need long-term monitoring, behavioural support, treatment adjustment and coordinated access to nutrition, physical activity guidance, pharmacotherapy and, in some cases, metabolic or bariatric surgery.

That mismatch has created a treatment gap. Specialist obesity services remain scarce, unevenly distributed and difficult to access, even in high-income countries. Waiting times, geographic barriers and workforce shortages mean many patients receive no structured care or only low-intensity support that is unlikely to produce sustained weight loss or meaningful improvement in obesity-related complications.

AI is being explored as a way to expand the reach of clinical expertise. In obesity care, its potential value lies in its ability to process large and complex data streams, including electronic health records, wearable device outputs, imaging, genomic profiles, behavioural records and environmental data. These systems could help identify risk before obesity becomes severe, classify patients more accurately, guide treatment intensity and support patients between medical visits.

The review organizes AI tools by clinical function. Some systems are designed for prediction, using electronic health records and cohort data to forecast obesity risk or complications. Others are built for stratification, combining imaging, biomarkers, genetics and lifestyle data to identify subtypes of obesity that may respond differently to treatment. A third group supports clinical decision-making by helping clinicians choose among lifestyle intervention, pharmacotherapy or surgery. Another fast-growing category includes AI-enabled behavioural coaching, where digital platforms provide personalized lifestyle prompts and feedback. Generative AI systems, including large language models, are also being tested for documentation, patient education and communication support.

Among these areas, early risk prediction appears to be one of the most mature. The review describes electronic health record-based models that can identify children at high risk of developing obesity within one to five years. These systems use data such as growth patterns, perinatal factors, comorbid conditions, medication histories and sociodemographic variables. Some paediatric datasets include millions of healthcare encounters, allowing models to detect patterns that would be difficult to identify through routine clinical review.

Adult risk prediction models also show promise. Studies using national health datasets have tested machine learning systems that assess obesity risk through metabolic variables such as triglycerides, liver enzymes and uric acid. Some models show different predictive patterns by sex and age, suggesting that obesity risk tools may need to be tailored to patient groups rather than applied as universal calculators.

Interpretability is a major issue. The review stresses that clinicians need to know why a model flagged a patient as high risk. Tools that explain the contribution of specific variables can turn AI predictions into practical clinical guidance. If a system identifies sedentary time, daily step count or fasting glucose as important drivers of risk, those outputs can support targeted counselling rather than producing an unexplained warning.

The strongest current evidence is not the same as proof of real-world benefit. Many risk prediction models have been tested only within the systems where they were developed. External validation across different health systems, countries, ethnic groups and socioeconomic settings remains limited. The review notes that no prospective implementation study has yet shown that embedding these tools into routine care reduces obesity incidence or improves long-term health outcomes.

AI is also being used to evaluate environmental and behavioural determinants of obesity. Models trained on satellite and street-level imagery can assess neighborhood features such as walkability, green space, road structure, building density and access to recreation. These tools may help public health agencies identify areas where built environments increase obesity risk and where interventions could be targeted.

This type of AI is more useful for public health planning than individual diagnosis. It can help governments decide where to invest in safer walking routes, green spaces or community health programmes. But the evidence still largely shows associations between environment and obesity prevalence, not proof that AI-guided urban planning reduces obesity rates.

AI could move treatment beyond BMI, but precision care remains unproven

AI could help obesity medicine move beyond body mass index as the main classification tool. BMI is simple, widely used and useful for population-level assessment, but it does not capture the biological complexity of obesity. People with the same BMI can have different patterns of visceral fat, muscle mass, metabolic dysfunction, inflammation, genetic risk, eating behaviour and treatment response.

AI-based phenotyping aims to identify those differences. Multimodal models can combine electronic health records, imaging, genetics, metabolomics, microbiome data, wearable signals and behavioural records. These systems may classify patients into more meaningful subgroups, such as those with high visceral adiposity, metabolic-inflammatory profiles, sarcopenic obesity or behavioural patterns linked to hedonic eating.

The potential clinical impact is significant. A patient with high visceral fat and insulin resistance may need earlier pharmacotherapy or closer cardiometabolic monitoring. A patient with muscle loss and excess fat may need a treatment plan focused on resistance training, nutrition and preservation of lean mass. A patient whose obesity is strongly linked to eating behaviour and cravings may require more intensive behavioural therapy. AI could eventually help match these patients to the right treatment pathway earlier.

The review finds that multimodal AI models often outperform models based on only one type of data. By combining genetic, clinical, imaging and lifestyle inputs, these systems may better predict obesity risk, cardiometabolic complications and treatment response. Large biobank studies and multimodal reviews suggest performance gains when models integrate multiple sources of information rather than relying on BMI or isolated biomarkers.

However, this is also one of the areas where the gap between scientific promise and clinical readiness is widest. Many phenotyping models can identify subgroups, but there is still little prospective evidence showing that using those AI-defined subgroups to guide treatment improves patient outcomes. In other words, a model may classify obesity more precisely, but clinicians still need proof that acting on that classification leads to better weight loss, better metabolic health, fewer complications or lower costs.

The review also identifies cost and access barriers. Imaging, genomics and multi-omics testing are not available to all patients or all health systems. If precision obesity care depends only on expensive data sources, it may widen existing disparities. The authors note that electronic health record-based deep phenotyping could offer a lower-cost alternative, using routinely collected data such as comorbidities, prescriptions and weight trajectories to identify clinically useful subtypes.

Equity remains a major concern. Many AI models are trained on data from high-income countries, digitally connected patients and majority populations. Genomic and multimodal datasets often underrepresent ethnic minorities, lower-income groups and people in low- and middle-income countries, even though obesity prevalence is rising rapidly in many of those settings. If AI systems are trained on narrow datasets, they may produce less accurate predictions for the very patients most in need of better care.

Clinical decision support tools raise another set of risks. AI could help clinicians decide when to escalate from lifestyle treatment to medication, when to refer for metabolic or bariatric surgery, or how to combine GLP-1 receptor agonist therapy with behavioural support. But the review warns that these systems must be integrated into clinical workflows with clear accountability. AI should help organize evidence and risk signals, not replace medical judgment.

GLP-1 receptor agonists have become a major part of obesity management, creating new demand for systems that can monitor adherence, guide nutrition, support muscle preservation and reduce the risk of weight regain after treatment changes. The review points to a growing model in which AI-enabled behavioural coaching is paired with medication to support longer-term outcomes.

This combination may become an important direction in obesity care. Medication can drive weight loss and metabolic improvement, while AI-enabled coaching can support lifestyle change, track behaviour and provide feedback between clinic visits. But the evidence for additive long-term benefit remains limited, and health systems will need to evaluate whether these combined models are effective, affordable and equitable.

Behavioural coaching shows the clearest patient benefit, but oversight is essential

AI-enabled behavioural coaching is one of the most clinically active areas reviewed. These systems use AI to deliver personalized lifestyle support through mobile apps, conversational agents, adaptive prompts, wearable feedback and digital coaching platforms. They can adjust the timing, frequency and tone of messages based on step counts, sleep patterns, heart rate, diet logs, mood reports and user engagement.

The reason this area matters is that obesity treatment depends heavily on sustained behaviour change, but health systems rarely have enough staff to provide continuous support. AI coaching could create a persistent layer of care between visits, helping patients maintain nutrition goals, increase physical activity, monitor progress and respond to setbacks.

The review finds emerging evidence of clinically meaningful weight loss from AI-enabled behavioural interventions. A randomized trial within a Diabetes Prevention Programme framework found that AI-powered lifestyle coaching produced outcomes comparable to human coaching among adults at high risk of type 2 diabetes, many of whom had overweight or obesity. This is one of the clearest signs that AI coaching could help scale behavioural care without entirely depending on a limited human coaching workforce.

Other evidence from smartphone-based nutritional interventions and telemedicine-supported care suggests that digital support can reduce weight, BMI and waist circumference. But the review cautions that app-only interventions without meaningful personalization or human support tend to show smaller effects. The strongest models may be hybrid systems that combine AI support with clinician oversight and structured care pathways.

The regulatory landscape is beginning to shift as AI-driven digital therapeutics and continuous glucose monitoring tools enter weight management. These systems can provide biomarker-informed lifestyle prompts, helping users understand how food choices, activity and daily habits relate to metabolic responses. The review treats these developments as a sign that AI in obesity care is moving from research into real-world products.

However, commercial deployment brings risk. A digital tool can reach patients quickly, but wide adoption before proper validation could lock in biased or unsafe systems. The review warns that many AI obesity tools lack long-term effectiveness data, cost-effectiveness evidence and robust post-deployment monitoring. Some studies are short, small or limited to digitally literate populations in high-income settings.

Safety concerns are even sharper for generative AI. Large language models may be useful for drafting patient education materials, summarizing information or helping communication, but they can produce inaccurate statements, unsupported recommendations or advice that conflicts with clinical guidelines. The review argues that patient-facing generative AI in obesity care should not operate without human oversight.

Human-in-the-loop governance is presented as a core requirement. In practical terms, AI-generated recommendations or education should be reviewed, edited and approved by qualified clinicians when clinical risk is meaningful. Professional accountability should remain with healthcare providers, not the software. AI should function as a clinical support layer, not an autonomous decision-maker.

Bias auditing is another urgent requirement. The review notes that fewer than 20 percent of reviewed studies reported performance metrics across demographic subgroups. That is a serious weakness in a disease area shaped by social inequality, food environments, income, ethnicity, geography and access to care. Without equity-stratified reporting, health systems cannot know whether AI tools work equally well for all patients.

Post-deployment surveillance is also necessary because AI systems can degrade over time. Patient populations change, clinical practice changes, data quality changes and user behaviour changes. A model that performs well at launch may become less accurate later. The review calls for ongoing monitoring of safety, outcomes, subgroup performance and new inequities after AI tools enter clinical use.

The authors also highlight sensitive populations, including pregnant women with obesity. AI-driven lifestyle interventions may help improve diet, physical activity and metabolic health during pregnancy, but the stakes are higher because poor advice or algorithmic error could affect both mother and child. Tools designed for such groups require strict validation, informed consent, data governance and clinician supervision.

For health systems, the next phase should focus on disciplined implementation. AI tools need external validation in local populations, clear reporting standards, bias audits, human oversight, regulatory alignment and long-term monitoring. For clinicians, AI may become a useful co-pilot, but it should not displace clinical judgment. For patients, the safest use of AI will be as part of multidisciplinary care, not as a stand-alone substitute.