The aim of this study was to find a clinical marker for identifying refractory cancer cachexia. We analyzed computed tomography imaging data, which included the third lumbar vertebra, from 94 patients who died of uterine cervix or corpus malignancy. The time between the date of examination and date of death was the most important attribute for this study, and the computed tomography images were classified into >3 months before death and ≤ 3 months before death. Psoas muscle mass index was defined as the left-right sum of the psoas muscle areas (cm2) at the level of third lumbar vertebra, divided by height squared (m2). A data set of 94 computed tomography images was obtained at baseline hospital visit, and a data set of 603 images was obtained at other times. One hundred (16.6%) of the 603 non-baseline images were scanned ≤3 months before death. Mean psoas muscle mass index change rates at >3 months before death and ≤3 months before death were -1.3 and -20.1%, respectively (P < 0.001). Receiver operating characteristic curve analysis yielded a cutoff value of -13.0%. The area under the curve reached a moderate accuracy level (0.777, 95% confidence interval 0.715-0.838). When we used the cutoff value to predict death within 3 months, sensitivity and specificity were 74.0 and 82.1%, respectively. Measuring change in psoas muscle mass index might be useful for predicting cancer mortality within 3 months. It could become a potential tool for identifying refractory cancer cachexia.