OBJECTIVES: This study aimed to review publications investigating the prediction of clinical outcomes, HCRU and costs using secondary databases. METHODS: A search syntax with keywords on prediction models and secondary data was run on 31/01/2020 on MEDLINE. Publications meeting the following criteria were included: English language, published in 2015-2019, use of secondary data (claims/electronic medical records [EMR]/ registries and alike), use of prediction model(s) and humans. Outcomes of interests were mortality, morbidity, socio-economic characteristics, prescriptions, HCRU and costs. Publications on other outcomes were excluded. Abstracts were reviewed by two researchers for data extraction. RESULTS: Out of 511 initial studies, 284 were excluded (66.5% study design and 31.7% outcome out of scope; 1.8% duplicate study). The remaining 227 studies presented predictions using claims data (56.8%), EMR (22.5%), administrative data (8.4%), multiple sources (5.7%), registries (5.3%) and other sources (1.3%). They largely used traditional parametric regressions and matching (n=169, 74.5%) and relatively rarely used (supervised) machine learning methods (n=29, 12.8%) for prediction. In 29 studies, information on methods was not available in abstracts. The top three indications investigated were cardiovascular diseases (n=39, 17.2%), oncology (n=29, 12.8%) and mental diseases (n=23, 10.1%). The most frequently predicted outcomes were HCRU (n=91, 40.1%), morbidity risk (n=73, 32.2%) and mortality (n=41, 18.1%). Costs were predicted relatively rarely (n=20, 8.8%), without a specific indication focus and with little information on prediction quality: Only in 5 studies, quality of cost predictions was tested; only in 1, it was quantified (via area under the receiver operating curve). CONCLUSIONS: Although predictions of HCRU and morbidity risk were common, cost outcomes, especially in “cost-intensive” indications, were relatively rare with very limited information on prediction quality. Addressing the evidence gaps in cost predictions can greatly inform payor decisions. Claims databases, as powerful sources of real-world costs, can be mobilized in this respect.