Advances in artificial intelligence empowering early-stage human health: Current landscape and future directions
While healthcare providers have used computer-aided programs since the 1950s, artificial intelligence (AI) in health promotion has only recently flourished, driven by advances in large language models (LLMs), discriminative machine learning, and multimodal foundation models. Fueled by growing data and computing power, AI now excels in drug development, diagnostic support, and AI-assisted surgery. The ancient Eastern concept of "preventing disease before it occurs" and modern preventive medicine both aim to safeguard health through early proactive intervention. Amid global aging and rising incidence of complex chronic diseases, which are often hard to treat once developed, early health promotion is increasingly crucial. Studies suggest AI tools hold great promise for reshaping early-stage health promotion, yet this field faces distinctive challenges, and its future development pathways warrant in-depth considerations.
Current landscape
Early precision prevention relies on identifying high-risk individuals. A recent study shows AI's remarkable potential in this area. Trained on data from 402,799 UK Biobank participants (including demographics, lifestyle, and 1,256 diseases) and validated on the 1.93 million Danish registry records, the Delphi-2M predicts over 1,000 diseases and mortality with mean area under the receiver operating characteristic curve (AUC) of 0.69 (UK) and 0.67 (Danish) over 20 years.1 However, the model has limitations, including healthy volunteer bias and the impact of diverse data sources and missingness patterns on predictions and variable performance across ancestry and deprivation groups. These issues highlight the need for caution with heterogeneous healthcare datasets and position the model as a supplement, not a replacement, for clinical judgment.
Early screening is key to preventing multisystem diseases from becoming chronic. AI has advanced significantly in addressing long-standing gaps in early disease screening. For example, using 12-lead electrocardiograms and basic demographics, EchoNext identifies a wide range of structural heart diseases with an AUC of 85% in internal validation and 78%–80% in external cohorts, outperforming 13 board-certified cardiologists (64% accuracy). Additionally, a “silent deployment” of EchoNext achieved a 74% positive predictive value for structural heart disease in individuals without prior echocardiograms, addressing the field's heavy reliance on this method. However, the small sample size of the prospective trial is a key limitation, and larger studies are needed to confirm benefits in routine care.
Maintaining a healthy lifestyle is a well-documented way to prevent diseases and prolong life. Promising evidence suggests a paradigm shift from passive data collection to active health guidance. For instance, personal health LLM (PH-LLM) outperformed human experts on sleep medicine (79% vs. 76%, n = 629) and fitness (88% vs. 71%, n = 99) multiple-choice exams sourced from BoardVitals and aligned with NSCA-CSCS standards, respectively. PH-LLM also performed similarly to experts in an evaluation involving 350 real-world case studies.3 However, exam scores are only surrogate measures reflecting knowledge proficiency rather than long-term health impact. Another limitation is sample representativeness bias (e.g., fitness cases enriched with males aged 30–59 years), weakening the generalizability.
AI-based virtual reality (VR) exercise systems have shown promise in promoting healthy behaviors.4 For example, REVERIE, an AI-based VR system using deep reinforcement learning to train transformer-based virtual coaches, effectively reduced the primary outcome of fat mass (mean −4.28 kg, vs. control) in adolescents with excess body weight.4 This effect did not differ significantly from real-world physical sports. However, the study did not assess the system's impact on cardiorespiratory fitness and lacked data from long-term follow-up data.
