F2O: Artificial Intelligence Predicts Erectile Function After Prostatectomy by Decoding Surgical Gestures

KEY FINDINGS

• The system: F2O (Frame-to-Outcome) is an end-to-end artificial intelligence system that converts raw surgical video of the nerve-sparing phase of robot-assisted radical prostatectomy into sequences of consecutive "surgical gestures" without human intervention.
• Gesture recognition performance: AUC 0.80 at frame level; AUC 0.81 at video level.
• Outcome prediction: The gesture features produced by the system (frequency, duration, transitions) predicted postoperative outcomes with an AUC of 0.79; this statistically overlapped with the 0.75 performance of a model based on manual human annotation.
• Agreement: Effect directions were concordant across 25 shared features (Δdort ≈ 0.07), correlation r = 0.96.
• Protective technical patterns: Two patterns associated with the recovery of erectile function were quantitatively established - prolonged blunt tissue peeling and reduced use of energy (cautery).
• Data source: The model was trained on data collected from 5 international centres; prospective external validation is planned.

Background: Why Could Surgery Not Be Measured?

In radical prostatectomy, nerve-sparing dissection is the most critical and most technique-dependent step, aiming to preserve erectile function without compromising oncological safety. Nevertheless, how this phase should be performed has for decades been transmitted largely through the master-apprentice tradition and experiential opinion. In the words of Professor Andrew J. Hung (Cedars-Sinai Medical Center): this operation has been learned in the way mentors taught it, through dogma; but it has not rested on a scientific basis.

The scale of the problem matters: in multicentre series, the rate of recovery of erectile function after nerve-sparing prostatectomy runs at approximately 35%. Moreover, because there is typically a one-year delay between the operation and the moment the functional outcome emerges, the surgeon lacks any meaningful feedback loop regarding their own technique.

Study Design: The "Alphabet" of Surgery

Hung and his team have developed a classification system that reduces a complex operation to discrete manoeuvres of approximately 2 seconds - "surgical gestures". These gestures cover basic actions such as cutting, pushing, cauterisation and retraction, along with blunt dissection, sharp dissection and combination manoeuvres. In earlier work the team had shown that these gesture sequences correlated with postoperative outcomes; but because second-by-second manual annotation of videos is labour-intensive, the method could not be scaled.

The technical innovation of F2O: The system combines transformer-based spatial and temporal modelling with frame-based classification. In raw video at 30 frames per second, each gesture corresponds to approximately 60 frames; the model determines for each frame which manoeuvre is being performed. The second critical innovation is change point detection: automatically determining when one gesture is completed and the next begins. This establishes an uninterrupted pipeline from raw video to a single outcome prediction that requires no human labour.

Key Results

Which manoeuvres are protective?

The most clinically striking output of the study is the first quantitative definition of the technical patterns associated with recovery of erectile function:

• Prolonged blunt tissue peeling: Separating the neurovascular bundle from the prostate with blunt manoeuvres along natural tissue planes is associated with the prevention of nerve injury.
• Minimising the use of energy: Even when nerves are anatomically preserved, thermal injury can lead to loss of function; reducing cautery use was found to be protective at the gesture level.

Although these two principles have long featured in surgical intuition, they have for the first time been placed on a reproducible and quantitative evidence base.

Implications for Clinical Practice

1. Closing the surgical feedback loop

The one-year gap between operation and functional outcome can be bridged retrospectively through artificial intelligence. Matching the gesture profile of a surgeon's own cases with the recovery probabilities of their own patients allows the manoeuvres to maintain and to avoid to be determined objectively.

2. Making training objective

The teaching of nerve-sparing technique can shift from subjective video assessment to measurable performance indicators. This carries the potential for structural transformation particularly in robotic surgery training programmes and competency certification.

3. Prediction in patient counselling

In the longer term, individual recovery probability estimates derived from the operative video could provide a concrete basis for expectation management with the patient and for planning penile rehabilitation.

4. Methodological limitations

This study is an enabling development that brings together existing components. Although the model was trained on data from five international centres, the real test of generalisability is independent external validation. The investigators are planning a prospective clinical study to test whether gesture sequences genuinely affect postoperative recovery. In addition, outcome assessment is limited to erectile function; oncological outcomes and continence fall outside the scope of this analysis.

Conclusion

The F2O system is a notable step towards moving surgery from being an "art" to a measurable science. In Professor Hung's analogy, just as the automated sequencing of the human genome revealed the genetic basis of countless diseases, decomposing a complex operation into consecutive gestures and analysing it in a scalable way carries the potential to reveal the elements that define "good surgery".

In robotic uro-oncological surgery, the decisive influence of technical detail on functional outcome is well known. The real contribution of this study is that it has provided the methodological infrastructure to move that influence from opinion to evidence. The transition into clinical practice depends on the results of prospective validation.

References

• Li X, Matsumoto N, Pasupulety U, et al. End to end AI system for surgical gesture sequence recognition and clinical outcome prediction. npj Digit Med. 2026;9(1):494. doi:10.1038/s41746-026-02927-5
• Hung AJ. AI system links surgical gestures with postoperative outcomes. Urology Times, 26 August 2026.
• arXiv:2511.11899. End to End AI System for Surgical Gesture Sequence Recognition and Clinical Outcome Prediction.
• Medical Xpress. AI analyzes surgical technique to improve prostate cancer care.