Voice-based depression assessment in patients with gynecological cancer.

161 Background: Although depression in patients with cancer can lead to a deterioration of treatment adherence and quality of life (QOL), diagnosis of depression in patients with cancer is difficult due to symptoms overlapping with side effects of cancer treatment such as fatigue. Traditional screening tools like the PHQ-9 questionnaire are underutilized in clinical practice. Recently, interest has grown in utilizing patients' voices to assess their mental well-being. This study aims to evaluate depression in patients with gynecological cancer and explore the usefulness of voice analysis as a depression detection tool. Methods: PHQ-9 scores were collected from 197 patients with gynecological cancer, with 28 cases undergoing longitudinal assessments at diagnosis, postoperatively, and during chemotherapy. Voice recordings were obtained from 70 patients at diagnosis, alongside serum samples from 22 patients. The PHQ-9 scores were compared across different treatment periods, with a specific focus on patients exhibiting a PHQ score of ≧10, who can be diagnosed with Major Depressive Disorder (MDD). The provision of supportive care for MDD patients was also verified. A random forest model was developed from voice features obtained from 70 patients, with hyperparameter tuning and cross-validation to predict mild depression. Serum metabolites were comprehensively analyzed between depression prediction group and normal prediction group by the depression prediction model. Results: Mean PHQ-9 scores decreased over time since diagnosis; initially 6.50 ± 4.99 (mean ±SD), post-operatively 6.18 ± 4.25, and during chemotherapy 4.88 ± 4.29. The frequencies of MDD also decreased over the clinical course; at diagnosis 24.3%, post-operatively 15.5%, and during chemotherapy 15.4%. Out of 17 cases of MDD during chemotherapy, only one (6.25%) received psychiatric intervention. All MDD cases during chemotherapy had PHQ-9 scores of 5 or higher, representing mild depression at diagnosis. Mild depression at diagnosis could be predicted with an AUC of 0.89 using voice features. Metabolites identified by the voice-based depression prediction model were associated with depression, such as xanthine, methionine, and taurocholic acid. Conclusions: Depression during chemotherapy is often not intervened. Patients with mild depression at diagnosis tend to develop MDD during chemotherapy. Voice-based mental health assessment at diagnosis is possibly a promising screening tool for depression during subsequent cancer treatment. Clinical trial information: UMIN000044266 .