Using voice recognition and machine learning techniques for detecting patient-reported outcomes from conversational voice in palliative care patients.
Lei Dong, Hideyuki Hirayama, XueJiao Zheng and 2 others
PMID 39778050WHAT IT FOUND
The symptom detection model for palliative care conversations was not ready for clinical use.
Patient speech was transcribed accurately only 55.6% of the time, and detection scores were low.
Key findings
01The model's primary performance measure, the F1 score, was 0.36 for pain, 0.46 for anxiety, 0.45 for depression, 0.43 for feeling at peace, and 0.31 for practical matters.
02In 100 palliative care interviews, the automatic transcription tool recognized 76.1% of all words, 55.6% of patient words, and 82.2% of healthcare provider words.
03Stroke patients had the lowest patient speech recognition rate at 48.3%, but recognition rates did not differ significantly across disease groups (p = .515).
STILL TO COME
How it was doneWhat they foundWhat it means for RNs
Read the rest of this summary
You get three full summaries a month, free, and we do not ask for a card. Search, the TL;DRs and your library stay unlimited either way.
What it does not show
The study was done in a home-visit palliative care setting, so it may not apply to palliative care units, clinics, or wards. The sample was small for training a machine-learning model. The model was trained on automatic transcriptions that were often inaccurate, especially for patient speech. The study used an auto-machine-learning black box, so its reasoning could not be adjusted or easily interpreted. Only an iPad was used, and tablet position, background noise, and distance may have reduced speech quality. The sample included many older adults, 23 patients with dementia, and stroke patients had the lowest recognition rate, which may limit general use. The study measured transcription and model performance, not patient care, nurse documentation burden, or clinical decisions.
Declared interests
The study was funded by the Japan Society for the Promotion of Science KAKENHI grant 22K11240, with no funder role in design, collection, analysis, writing, or submission. The authors declared no conflicts of interest.
The easy way to misread this
Do not conclude that voice recognition and machine learning can replace nurse assessment of palliative care symptoms. The model's F1 scores were 0.36, 0.46, 0.45, 0.43, and 0.31, and patient speech was transcribed accurately only 55.6% of the time.