- 195 → 120 s
- median consultation, without vs with PIA
- ≈ 37 min
- physician time saved per day at 30 CT patients
- 3 cities
- with follow-on projects: Leipzig, Berlin, Munich
The study at a glance
Site
TUM Klinikum rechts der Isar in Munich, in routine CT operations.
Design
Prospective and comparative, approved by the institutional review board.
Patients
86 patients scheduled for CT: 43 prepared with PIA, 43 with the physician consultation alone.
Setting
Pre-CT informed consent on an iPad, with text-to-speech for patients who prefer to listen.
Measured
Consultation time, patient comprehension, clarity and satisfaction, plus expert ratings of every conversation in five categories.
Published
Journal of Imaging Informatics in Medicine (Springer), 2025.
The challenge
Informed consent before medical imaging is essential but consumes a disproportionate amount of physician time. Studies show consent counseling can take over 11 minutes per patient, roughly 7% of a radiologist’s professional time. The clinical need is clear: patients deserve thorough information, but radiologists should spend their time on diagnosis and treatment, not repeating standardized information.
Hospital-information deployment; the published study covered pre-CT informed consent.
What we built
PIA (Patient Information Assistant) is a RAG chatbot that delivers pre-CT informed consent information to patients before the physician consultation, running on an iPad with text-to-speech. Clinical consent documents are split into paragraphs and embedded using a German-language model stored in a vector database. When a patient asks a question, the query is matched against the knowledge base and the most relevant documents are retrieved and supplied to the language model, which generates a grounded answer. Answers are grounded in validated clinical documents, which reduces the hallucination risk that makes standalone LLMs unsuitable for clinical use.
More about PIA: pia-chat.com
The study
We validated PIA in a prospective comparative study with 86 patients scheduled for CT imaging at TUM Klinikum rechts der Isar in Munich. Patients were randomly assigned to receive pre-CT information either from PIA (n=43) or through the standard physician consultation alone (n=43). The study was approved by the institutional review board and published in the Journal of Imaging Informatics in Medicine (Springer, 2025).
Consultation time
| PIA group | Control group | ||
|---|---|---|---|
| Median consultation time | 120 seconds | 195 seconds | p = 0.04 |
| IQR (Interquartile Range) | 100–140 s | 170–220 s |
Patients who used PIA before seeing the physician required 38% less consultation time, a statistically significant reduction.
Patient comprehension and satisfaction
| Metric | PIA | Control | p-value |
|---|---|---|---|
| Information clarity | 8.64 ± 1.69 | 8.86 ± 1.28 | 0.82 |
| Completeness of information | 8.67 ± 1.83 | 8.81 ± 1.87 | 0.58 |
| Overall comprehension | 8.81 ± 1.40 | 8.93 ± 1.61 | 0.35 |
| Satisfaction with the delivery method | 8.88 ± 1.43 | 9.16 ± 1.43 | 0.14 |
Both groups rated clarity, completeness, comprehension and satisfaction similarly high; none of the differences was significant. Consultation time falls, with no significant difference in comprehension.
Clinical quality
Two independent radiologists rated all 43 PIA conversations across five categories: overall quality, scientific evidence, clinical usefulness, consistency, and up-to-dateness. Both rated the overall quality of every conversation very good or excellent, and every conversation as evidence-based and up to date.
Where physicians remain superior
The study also showed that physicians were significantly more effective at alleviating emotional concerns about the upcoming examination (8.30 vs 6.46, p=0.003). This confirms the intended design: PIA handles information delivery, freeing the physician to address worries and build trust.
For a radiology department processing 30 CT patients per day, PIA translates to roughly 37 minutes of physician time saved per day, with patients arriving at the consent discussion already informed. Projects in obstetrics and surgical informed consent are underway in Leipzig, Berlin, and Munich.
Busch, F., Kaibel, L., Nguyen, H. et al. Evaluation of a Retrieval-Augmented Generation-Powered Chatbot for Pre-CT Informed Consent: a Prospective Comparative Study. Journal of Imaging Informatics in Medicine 38, 4312–4323 (2025).doi.org/10.1007/s10278-025-01483-wRead the full paper