Hi everyone,
I would like to develop and test a surgical scheduling decision support system, to predict rates of surgical delays and cancellations. I plan to collect retrospective data on patient demographics, health history, surgical times, etc to train the model. The prototype will then be tested against prospective data for sensitivity & specificity.
How can i go about this?
First you find the data (or a suitable dataset), then you clean the data.
Perhaps at this point you want to see which parts of data is most useful to you and use that only to predict your output! That means that you have to perform feature analysis, check Kaggle courses…
Next you build an ML model (may be able to use a lot of frameworks depending also on type of ML model), perhaps for tabular data the tree and forest models seem to perform well.
Then you train this model with the data you gathered, and fix parameters as they come along according to the evaluation of the training phase.
And hopefully once you get good performance on the previous stage then you have a good model to be used to predict.
Thats a rough guideline!
To the previous sensible reply I suggest also
-
survey what is already being discussed in this speciality, such as Fujitsu's quantum-inspired computing helps uncover additional OR capacity | Healthcare IT News
-
think broadly about what optimization means here; it’s potentially a very complex topology.
Operating Room Performance Optimization Metrics: a Systematic Review - PMC
To support what you suggest about ‘optimization’ in this case there are also economic and social ‘externalities’ to consider.
I have a good older friend who he and his wife are reasonably well known regional artists-- Though their son became a big-wig ME atIntuitive / Da Vinci Surgical (i.e. one of the robots).
I should remember to ask him next time we speak where they are at with all this.
Indeed. In fact, each set of stakeholders - surgeons, nursing staff, hospital CFO, patients, etc - may have different and competing priorities and success metrics.
Unfortunately, not only the people building such models, but more importantly those willing to pay for them (unless this Bhutan, and I’m not sure that ‘heard factoid’ is still true) are optimizing for ‘overall happiness’ (and then ).
Thank you! I plan on taking into account nursing resource allocation (their years of experience, skillset, etc) into my model.
My takeaway from @Nevermnd ’s reply is that some important components of ‘optimization’ of a clinical service may not be limited to dollars and hours quantitative factors. Patient and professional staff satisfaction, for example. I think this is mentioned in the survey paper linked above. To paraphrase Frank Herbert, ‘Label depends on viewpoint…[don’t] … look at the universe through too narrow an opening.’
Keep us posted
hey @Mathumesa
actually that’s a great idea to address treatment delays but as I see you planning to take nursing resource allocation, I am not sure how you will find this resourceful as addressing surgery or operating room depends on surgeon, nursing staff, how emergency is the treatment, life expectancy if surgically treated, OTs available, OT condition, Surgical material available, and most importantly patients data.
Also remember Healthcare data comes with its own privacy and protection laws (HIPAA and PHI)
Now sharing points something which might be useful, if you have resources and can get MIMIC III (Database regarding patients icu addmissions) which when used for research, patient’s identification information is hidden in these database, so this can be useful for your ml model.
Next don’t forget to incorporate ABCs of treatment emergencies which again will help you categories your reason of delay in surgery or any concomitant setbacks.
Nursing resources will surely be helpful depending on what it contains like patient case history details (without personal identification), images, lab, diagnosis, treatment advised (this you can again divide into surgical and non-surgical, so you can remove patient data with non-surgical data as you are addressing only surgery delays but use non surgical data as confounding factor as some medication intake cause inability to operate on patients, like patient on blood thinner, patient with high BP and high diabetic level, patient with active infection, patient on chemotherapy(for cancer treatment) etcetera.
Hope this helps create your idea!
Regards
DP