1 of 13

BAUS Oncology Meeting 2024�East Midlands Conference Centre, Nottingham

The RALPulator.

Using Machine Learning and Natural Language Processing to Predict the Operative Times of Robotic Assisted Laparoscopic Prostatectomy

J. Wilson¹, R. Lee³, H. McDonald⁴, A. Gomati², J. Eaton¹, B. Patel¹, A. Okeke¹, M. Aktar², J. Peacock ¹, E. Tudor¹

Cheltenham General Hospital, Gloucestershire Hospitals NHS Foundation Trust
Hereford General Hospital, Wye Valley NHS Foundation Trust
NHS North of England Commissioning Support Unit
Lancashire & South Cumbria NHS Foundation Trust

2 of 13

Background

>7.4 million on RTT waiting list
4.5 million pre pandemic
Urology- 418,000 (April 24)
Urology 285,000 �(2021)

3 of 13

Background

NHS Elective Recovery Plan
Elective Recovery High Volume Low Complexity (HVLC)
Independent Investigation of the NHS

4 of 13

Background

5 of 13

Methods

Retrospective analysis of 124 RALPS
Paper notes + electronic notes
Outcome = dependent variable = operative time (skin to skin)
Features = independent variables:

Surgeon, presence of trainee, nerve spare, PSA, MRI stage, prostate size, comorbidities, BMI, previous surgery, Gleason score, age

What is a RALP?!

6 of 13

7 of 13

Methods

Retrospective analysis of 124 RALPS (350)
Paper notes + electronic notes
Outcome = dependent variable = operative time (skin to skin)
Features = independent variables:

Surgeon, presence of trainee, nerve spare, PSA, MRI stage, prostate size, comorbidities, BMI, previous surgery, Gleason score, age

8 of 13

Methods

Using Python 3, creased a linear regression model (Sklearn)
Forward Feature Selection to optimise model
Ended up with:

Age, surgeon, trainee, stage, prostate size

Then created a linear regression model using a test train split of 75-25 i.e. we trained the model of 75% of the data and used the remaining, unseen data to evaluate the performance of the model.

We optomised to model by trying different combinations of features to get the best prediction overall. We settled on Age, surgeon, trainee, stage and prostate size as features the model used.

Note that some of the variable I thought would be useful were not included in the optimised model. For example, a nerve spare adds more time. However, multiple linear regression assumes all the variables are independent. I found one surgeon would do this procedure more often, and happened to be the quickest. Given these two features were correlated this meant adding this to the model reduced the overall accuracy of the model. ��I think a tree based model would be better to account for this.

9 of 13

Natural Language Processing
Pick out the relevant features from letters
Automatically plug these variables into the model
Give prediction
Create an interactive app

Methods

10 of 13

Results

https://ralpulator.streamlit.app

11 of 13

R2 of 0.70
Explains 70% of the variability in operative time as compared to “the mean model”

Results

https://ralpulator.streamlit.app

12 of 13

Results

Feature	Coefficient (minutes)	P-value	95% Confidence Interval
Constant	225.73	0.000	(159.06 - 292.41)
Surgeon 2	-60.26	0.000	(-75.95 - -44.56)
Surgeon 3	-34.51	0.003	(-56.82 - -12.19)
Trainee Present	27.00	0.040	(1.28 - 52.71)
Surgeon 1	14.57	0.176	(-6.71 - -35.90)
MRI – T3a	-12.38	0.155	(-29.54 - 4.79)
MRI – T2a	-8.39	0.330	(-25.47 - 8.69)
MRI – T2b	4.39	0.688	(-17.31 - 26.08)
Age at surgery	-1.04	0.056	(-2.10 - 0.03)
Prostate size (grams)	0.42	0.003	(0.14 - 0.69)

https://ralpulator.streamlit.app

13 of 13

Using local data we can better predict operative times
Can be used to inform theatre scheduling decisions and improve efficiency
Limitations- need more data
NLP needs proper training

Conclusions

https://ralpulator.streamlit.app