ERICA prediction

[edvardlindelof]

Stuff to do/ideas I want to try:

• [ ] QLasso implemented as in article. This is nice because it represents a benchmark as a thought-through model, and we can compare results to know if our implementation is working. https://web.stanford.edu/~bayati/papers/edwait.pdf
• [ ] LSTM and/or some other recurrent network. Since these networks have "memory" it would be interesting if one can feed the events more or less directly to the models. Are said to require huge amount of data for training but we might have enough.
• [ ] Nonlinear autoregression. Works well for multivariate time series prediction of electric load here http://cs229.stanford.edu/proj2016/report/GaoWuLiu-WindPowerAndElectricLoadForecasting-report.pdf
• [ ] Gradient boosting. Works well for multivariate time series prediction of sales here http://cs229.stanford.edu/proj2015/218_report.pdf
• [ ] Some feature reduction method. QLasso has this built in. Haven't read about this in the context of multivariate time series yet.

[edvardlindelof]

https://github.com/edvardlindelof/ERICA-prediction

[edvardlindelof]

QLasso feature engineering notes

0809: categorical time-of-week feature is quite important in article (e.g. for one hospital using only time-of-week feature gives mse of 2700 while including all features gives 1700). I've gotten the training mse down only to 8000 but I think this is about NAL and not about the method 0811: time-of-week only gives training mse 8200 (testing mse 530-2700 in article). Including low-prio-patients-inqueue feature reduces it by 400 (300, 500 etc in article). Including not-so-detailed workload features reduces it by 100 (100, 700 etc in article). Including a bunch of rolling averages of the wait times reduces it by 100 (20ish in article). Including the not-so-detailed workload/capacity features reduces it by 35 (almost 0 in article)

ToDo improvements

• changing the data aggregation to write data points only when low-prio patients arrive
• much more detailed workload features e.g. number of mep-patients that have received triage but not met doctor