Our study presents a detection model predicting a diagnosis of jaundice (clinical jaundice and occult jaundice) trained on prospective cohort data from slit-lamp photos and smartphone photos, demonstrating the model's validity and assisting clinical workers in identifying patient underlying hepatobiliary diseases.

This study demonstrated that deep learning models could detect jaundice using ocular images in blood levels with reasonable accuracy, providing a non-invasive method for jaundice detection and recognition. This algorithm can assist clinical surgeons with daily follow-up visits and provide referral advice. It also highlights the algorithm's potential smartphone application in sizeable real-world population-based disease-detecting or telemedicine programs.

Inclusion Criteria:

• The quality of slit-lamp images should be clinical acceptable. More than 90% of the slit-lamp image area, including three central regions (sclera, pupil, and lens) are easy to read and discriminate.

Exclusion Criteria:

• Images with light leakage (>10% of the area), spots from lens flares or stains, and overexposure were excluded from further analysis

This prospective, multicentre, observational study was led by Zhongshan Ophthalmic Centre (ZOC), Sun Yat-sen University, and conducted in three phases to collect data from participants from three surgery departments and three medical examination centres, including the Department of Hepatobiliary Surgery of the Third Affiliated Hospital of Sun Yat-sen University (HTH; Guangzhou, China), the Department of Infectious Diseases, Third Affiliated Hospital of Sun Yat-sen University (ITH; Guangzhou, China), the Department of Infectious Diseases, the Affiliated Huadu Hospital of Southern Medical University (HDH; Guangzhou, China), the Medical Centre of the Third Affiliated Hospital of Sun Yat-sen University(MCH; Guangzhou, China), Nantian Medical Centre of Aikang Health Care (NMC), and Huanshidong Medical Centre of Aikang Health Care (HMC).