Dengue virus is the most widespread arthropod-borne virus affecting humans, with as many as 528 million annual infections each year. Of particular concern are the subset of cases which develop into life-threatening dengue hemorrhagic fever, and those which further progress into dengue shock syndrome. Non-invasive tools that accurately differentiate dengue and its subtypes from other viral infections early in the disease progression are vital for timely therapeutic intervention and supportive care. Unfortunately, such tools are sorely lacking. Using liquid chromatography-mass spectrometry (LC-MS), we detect tens of thousands of molecular features in serum, saliva, and urine of suspected dengue patients in Nicaragua. We then use machine-learning methods to help identify candidate small molecule biomarkers which, along with easily obtainable clinical data, predict dengue diagnosis and prognosis. Our findings should aid in developing a low-cost diagnostic tool for use in the field.