LLM-assisted preoperative communication reduces prostate cancer patient anxiety

An LLM answered questions before the surgeon did

The study asked a practical question: if patients get personalized answers to their own preoperative questions before the standard face-to-face conversation, does that reduce anxiety and ease the physician's burden? Patients were newly diagnosed with prostate cancer and scheduled for radical prostatectomy at a single center, aged 18–80, with smartphone access and at least lower secondary education.

Randomization happened by hospital room rather than by patient, because rooms held three to four patients and the investigators wanted to limit conversation between groups. That design choice matters for reading the statistics: the analysis used room-level clustering, and the reported intracluster correlation coefficient was 0.03. Of 371 patients assessed, 276 were randomized and 268 were analyzed (130 control, 138 AI-assisted).

Both groups submitted questions through the same system. Only the AI-assisted group saw answers, delivered one hour after submission, generated by a locally deployed DeepSeek-R1 model drawing on a curated urology knowledge base. A content reviewer checked responses before delivery. All patients then had routine one-on-one counseling with one of three surgeons who were blinded to group allocation.

Anxiety and workload were the two primary endpoints

The trial had co-primary endpoints: patient anxiety after the final routine communication, measured by the GAD-7 scale, and physician workload, measured by the NASA Task Load Index (NASA-TLX). Because the standard GAD-7 asks about the past two weeks, the team used a modified 24-hour version to capture short-term change; the paper refers to both as GAD-7.

Final GAD-7 scores were 3.2 (SD 2.7) in the AI-assisted group versus 5.7 (SD 3.1) in controls. A linear mixed-effects model adjusting for baseline GAD-7 and physician, with a room-level random intercept, estimated an adjusted mean difference of −2.2 points (95% CI −3.4 to −1.1; P = 0.001). That 2.0-point gap was the difference the sample size was powered to detect as clinically meaningful.

Physician workload fell in parallel: NASA-TLX 39.9 (SD 16.9) versus 56.8 (SD 19.5), with an adjusted mean difference of −10.6 (95% CI −14.3 to −6.9; P < 0.001). Subscale reductions appeared in mental demand, temporal demand, effort, and frustration, but not in physical demand or performance. Baseline workload scores were not collected, so the workload comparison rests on final scores alone.

A second anxiety scale pointed the same way

The Visual Analog Scale for Anxiety (VAS-A), prespecified as a supportive measure of immediate anxiety, gave 2.0 (SD 1.5) in the AI-assisted group versus 3.8 (SD 1.8) in controls (P < 0.001). Anxiety severity categories also shifted toward lower levels in the AI-assisted group at final assessment, while baseline distributions did not differ between groups.

These are consistent signals, not independent confirmations of the same claim. VAS-A and the severity categories were not adjusted for multiplicity, and the paper labels secondary and exploratory analyses as supportive or hypothesis-generating. The GAD-7 result remains the one the trial was designed and powered around.

Emotion, illness perception, satisfaction, and time all moved

Secondary endpoints favored the AI-assisted group. Positive affect on the I-PANAS-SF was 11.3 (SD 3.0) versus 9.2 (SD 2.7), and negative affect was 10.2 (SD 3.6) versus 12.3 (SD 4.1). Illness perception on the B-IPQ was 27.5 (SD 13.2) versus 38.2 (SD 14.8), and satisfaction on the PSQ-18 was 57.0 (SD 9.0) versus 46.8 (SD 9.5).

Communication time dropped from 19.9 (SD 6.1) to 11.3 (SD 4.5) minutes overall, and the reduction held for each of the three physicians separately. Because these endpoints were not corrected for multiple comparisons, they should be read as a consistent pattern rather than as separately confirmed effects.

3,507 questions show the size of the information gap

Across both groups, patients submitted 3,507 questions, a mean of 13.08 per patient (range 1–39). The AI system answered 2,847 (81.2%) directly; the remaining 660 (18.8%) were judged unsuitable for a direct AI response and were answered from standardized templates by a content reviewer.

The most common questions were about urinary continence recovery (221 patients, 82.46%), the exact start time of surgery (208, 77.61%), and expected surgical duration (201, 75.00%). The authors read this volume as evidence of a substantial unmet informational need, and they invoke uncertainty-reduction theory to argue that unresolved information gaps can amplify anticipatory anxiety before major procedures. That mechanism is the authors' interpretation, not something the trial measured directly.

The AI prepared patients rather than replacing the surgeon

A key design feature is that the AI did not stand in for the physician conversation. Every patient still had routine face-to-face counseling with a blinded surgeon. The AI step came first, handling routine and frequently asked questions so that the in-person discussion could focus on individualized decision-making.

The authors propose this sequencing as the explanation for why patient psychological outcomes and physician workload improved at the same time: information was moved earlier and partly off the physician's plate. The trial supports the association between the AI step and the outcomes, but it does not isolate which component of the AI interaction produced the effect.

Within-group change is not the same as the trial's comparison

The AI-assisted group also showed significant within-group improvements in GAD-7, VAS-A, positive and negative affect, and B-IPQ after receiving AI answers, before the physician conversation. These exploratory analyses describe what happened inside one arm over time.

They cannot be substituted for the between-group comparison, because the control group did not receive an equivalent interim assessment of the same kind. Readers comparing the two sets of numbers should keep the denominators and comparison structures distinct: one is a within-patient before-and-after change in 138 patients, the other is a randomized between-arm difference.

Single-center results with a human in the loop

Several boundaries qualify how far these results travel. The trial ran at one institution, Fudan University Shanghai Cancer Center, between July and December 2025, in patients who had smartphone access and at least lower secondary education, so the findings do not automatically extend to patients with lower digital access or literacy. The authors themselves note that performance across different languages and cultural contexts still requires validation.

The human review step is part of the intervention as tested: nearly one in five questions needed a reviewer-written response, so the workload saving was achieved with a reviewer in the loop, not with a fully autonomous system. Physicians were blinded but could not be prevented from inferring allocation, so performance bias cannot be fully excluded.

The most decision-relevant reading is a comparison of where the benefit landed. The largest absolute gains were in physician-facing measures — roughly 8.6 fewer minutes per patient and about 17 points on NASA-TLX — while the patient-facing anxiety difference was about 2.2 GAD-7 points, right at the trial's prespecified clinically meaningful threshold. For a workflow deciding whether to add an LLM step before prostatectomy counseling, the evidence is strongest for time and workload relief with a reviewed system, and adequate but threshold-level for anxiety reduction.

Sources used in this answer