OBJECTIVES:  Patient registries play an important role in obtaining real-world evidence of the cost-effectiveness of treatments. However, their implementation is costly and sometimes infeasible in many middle-income countries (MICs). We explored the combination of data-mining and a large claims database to estimate the direct medical costs of HER2-positive breast cancer (BC) treatment in Iran and the fraction of total costs from trastuzumab use. METHODS:  Data from 21/03/2011-20/03/2014 were examined using data-mining to determine clinical stages. R-based classification algorithms were designed based on medication patterns including drug combinations and/or sequences of drugs in respect of three index dates (early, loco-regional and advanced BC). These algorithms were then validated using patient dossiers from the Cancer Institute of Iran and used to analyze claims data from the Iran Social Security Organization, a health insurer which covers approximately 50% of all Iranians (~40 million). A healthcare perspective was used to calculate the absolute and relative costs of medical services. RESULTS:  With an 84% (43/51) accuracy rate, 1,295 patients were identified and divided into three BC stages (early (n=802), loco-regional (n=125), advanced (n=218); some patients were categorized into >1 stage (n=177), none (n=284) due to insufficient information, while some were excluded (n=54) due to incomplete follow-up). Mean age per stage was 45, 46 and 48 years, respectively, while mean follow-up in all stages was approximately one year. Average costs of direct medical care in early, loco-regional and advanced stages were €11,796 (95%CI:€9,356-€12,498), €8,253 (95%CI:€6,843-€10,002) and €17,742 (95%CI:€15,720-€19,505), respectively. CONCLUSIONS:  When comprehensive patient registries are infeasible or costly, validated data-mining algorithms can support real-world cost-effectiveness analyses in MICs and thereby help to optimize reimbursement decisions. The accuracy of data-mining would be improved, if electronic reports of diagnostic tests were generated and included in the classification algorithms as well.