Abstract Background Epilepsy is one of the most common neurological disorders, affecting 1.1% adults in United States. Valproic acid (VPA) is an antiepileptic drug with narrow therapeutic window that requires close monitoring of the blood concentration. VPA is highly protein-bound, and the unbound free VPA (fVPA) is the bioactive form responsible for the pharmaceutical effects as well as the toxicity. However, fVPA is not commonly measured in most laboratories but is a test that needs to be sent out to reference laboratories, and the sendout test has a long turnaround time which delays timely patient care. The aim of the study was to develop fVPA estimating equations, to decrease the turnaround time and reduce adverse events, and ultimately improve patients’ outcomes. Methods We retrospectively collected the laboratory results of total valproic acid (tVPA), fVPA and serum albumin along with the patients’ demographic characteristics and other clinical data from 2013 to 2023 in our institution. While the albumin measurement was divided into normal (≥3.5 g/dl) and hypoalbuminemia (<3.5 g/dl) groups, the tVPA and fVPA were categorized into subtherapeutic, therapeutic, and supratherapeutic ranges. Using stepwise model selection, several linear regression equations were developed to predict fVPA concentration for total patients, patients with normal albumin only, and patients with hypoalbuminemia only. The estimated fVPA was then compared to the measured fVPA. The data were analyzed using SAS 9.4, SPSS 28.0 and EP Evaluator EE12. Results A total of 344 concurrent measurements of tVPA/fVPA/serum albumin from 157 patients were obtained, including 180 measurements from patients with normal albumin levels and 164 measurements from patients with hypoalbuminemia. In the patients with normal albumin levels, 32 (17%) were discordant between tVPA and fVPA. More than half of them had therapeutic tVPA but supratherapeutic fVPA. In contrast, 120 (73%) were discordant for the patients with hypoalbuminemia. About half of them had subtherapeutic tVPA but therapeutic fVPA. Several laboratory and clinical parameters including tVPA, albumin, age, sex, congestive heart failure, obesity, and diabetes mellitus were chosen in the final model. In all patients, our equation demonstrated a concordance rate of 66.3% and a bias of 1.46% between the estimated fVPA and the measured fVPA. Likewise, our equation displayed a concordance rate of 70.0% and a bias of -0.03% for patients with normal albumin levels. For patients with hypoalbuminemia, the concordance rate was 69.5% with a bias of 1.97%. In comparison to three previously published equations, our equations obtained increased concordance and decreased bias between estimated fVPA and measured fVPA. Conclusions Our data indicate that free VPA and total VPA do not correlate well according to the therapeutic ranges, especially in patients with hypoalbuminemia. Compared with previously reported equations, our equations showed enhanced concordance and reduced bias between estimated fVPA and measured fVPA, in both patients with normal albumin levels and patients with hypoalbuminemia. Further validation is needed to evaluate the accuracy and generalizability of the equations before clinical application.