Help with choosing a model

#1
Hello,

I apologize if I am posting in the incorrect place. If so, could someone please direct me or move this thread? Or if something similar has been asked can someone link me to the thread?

I am an accountant/finance guy and my knowledge of statistics and quant analysis is limited to my schoolwork many years ago as I don't use it much anymore. This is not work related, yet personal in nature. I have searched the web looking for anything I can find, but have come up empty or gone down the rabbit hole of links with nothing fruitful to take away. I figured I'd ask someone with a background in this type of analysis.

Problem: My son is 2-years-old and has what the doctors believe is a periodic fever syndrome. We went through the ringers for months as they have tried to figure him out like an episode of House. We are still waiting on some DNA test results to narrow things down, but the majority of life-threatening diseases have been ruled out during multiple hospital visits.

Being a parent, you can imagine I look for anything to fill my time trying to figure things out whilst not overstepping boundaries into the medical side of things where I am no doubt ill-equipped, although I now know way more about fevers than I ever thought I would.

I'll stop rambling and get to the point. He has episodes of high fever that are usually every 4 - 6 weeks. His temperature during non-episode periods are like any other person, 98.4 - 98.6 F. During episodes it usually spikes between 104.5 - 105.5 F, the max observed at 106.7. These will last anywhere from 3 - 7 days. Usually around 5 days. Sometimes his temperature will dip to 96.5 - 97.5 and this usually indicates he is about to spike to a high temperature. When I say "usually indicates" that is merely a subjective observation and has not been rigorously tested nor has it been measured in any type of quantitative environment. His temperature is always measured in his ear with the same thermometer (highest rated on market) that seems very consistent when taken right before doc appointments.

Data: I have a data set of the episodes dating back to 5/14 that include date, time, temperature, and whether medication was given (alternating ibuprofen and tylenol). I'm not sure if the medication variable is needed, but it does have a direct impact on the temperature as would be expected.

Problem: My goal is to be able to predict (within a few days) when the episodes might occur. I'd like to know this for multiple reasons: planning around my (and wife's) job, making sure he's drinking plenty of fluids (those high fevers can dehydrate a small child quickly), make sure he's eating as much as possible (he loses weight during episodes), giving the doctors as much helpful information as possible, etc.

There is also a treatment available at the beginning of an episode where a dose of steroids can be given and in some patients will knock out the fevers. We have not yet tried this but have the prescription from the rheumatologist. The drawback is that the episodes tend to come back faster. Also, if he has a normal infection (he still gets sick like a normal child which can blur the lines) and it's not an episode, the steroids weaken the immune system and can fuel the infection. I've tossed these out of the data set I currently have, but retained all of the information elsewhere.

So, predicting when the next episode might occur and identifying it as an episode quickly can possibly be vital to his treatment. I've tried to brush up on my stats and my old quant books, but any help with at least pointing me in the right direction for models to look at or any suggested techniques would be most helpful (I'm very proficient in Excel, but open to software suggestions or plugins). I can't thank you enough for taking the time to help me and my family with this. It will help fill my time and at least give me a sense that I am doing something other than reacting to symptoms.

I haven't attempted to look at any blood work data (during episodes his inflammation markers increase). He's had multiple labs drawn with each episode. If anyone has ideas for incorporating that, I'd love to hear them as well.

If I can figure something out that works well, I'm hoping to possibly create a little workbook that I can distribute to other families whose kids suffer from the same ailment so that they can just plug in numbers. Even if it's just to help them gear up for an episode.

Once again, thank you so much for your time and expertise. It means a lot.

Michael
 

rogojel

TS Contributor
#2
hi Michael,
i truly understand your problem and I find the idea to build a forevast model admirable. However, I must warn you that the first rule of all forecasting is that : " all forecasts are wrong" and if you trust a model too much you might put your child at higher risk, because you might assume that a new episode is unlikely when it would still happen.

That said, looking at the data would probably interest many people in this forum. I am not a specialist in time series models, so probably not the best person to do this, but if you can show your data I would be happy to look at it.

Others with more time series experience might chime in as well, as long as you are aware of the risk.

regards and good luck
 
#3
Thank you so much for your response. I was actually looking at time series models, so now I don't feel like I wasted too much time (and a little boost of confidence in my research skills at least). And yes, I definitely understand that the forecast might not be accurate, especially with the limited data I have now. I would never rely on it specifically, only use it as a tool to make the best guess possible for planning purposes and possibly to track which medication seems favorable. There are other factors and symptoms that are probably better indicators of when an episode may occur and once he's older and can communicate more efficiently it will definitely make things easier. When I get home I'll make a copy of my data and upload it for anyone who wishes to take a look. I currently have personal medical info on the spreadsheet I'm working off of, so I'll need to remove it. If I did find a model that seemed to predict when the episodes would come on quite well I would definitely give a deep explanation and disclaimer that it isn't to be relied on but merely a helpful tool. Thanks so much for the concern, I should have addressed that initially. I'm definitely digging this forum so far and will return for other endeavors, not that I can be much help. I enjoy fooling around with sabermetrics as a hobby. I'm repeating myself now, but thanks again! It means a lot.

Michael
 

CB

Super Moderator
#4
I really empathise with your situation and think it's great that you're wanting to approach this challenge using as much data and knowledge as you can.

But I do have reservations about whether developing this model independently is a good idea. Statistics are useful, but ultimately substantive knowledge about the topic is really important for developing good models. So my main suggestion would be to talk to your son's doctors about this. I would assume that if episodes like this are predictable with any reasonable accuracy, that they will have important background knowledge about how such prediction can best be completed.
 

Log33

New Member
#5
Again, thank you. I appreciate the warnings and reservations. I know people sometimes say that, but I really and truly do. It's given me second thoughts to ever distributing something out to anyone else who may rely on it for actual medical treatments. I will definitely talk to his doctors (I have before as well) and they are absolutely wonderful men and women. They're more interested (as they should be) with just making sure that he fits the criteria and that he's receiving the proper treatment. That they aren't overlooking something. But they field all of my questions with a smile (and there are many!).

If anyone feels uncomfortable, I completely understand. If anyone wants to have a look, I'll upload the data table with this post. We have an appointment with the rheumatologist next week, so I'll definitely be asking questions then. Honestly, the most likely scenarios for me to ever use this if it can get within a few days would be:

a) I tell my boss, ok these three days are the most likely when I'll need to be out, but not guaranteed. Then I'll probably work some overtime beforehand just in case, so little will need to be covered while I'm gone. Worst case scenario, I work a few extra hours if no fevers (which is a good thing).

b) When my son is older I can use it to give him a little lesson in statistics with something he can relate to and show him how we tried to predict when his fevers would come. I would have found that interesting. He might find it completely and utterly boring and want to go listen to Led Zeppelin records. Either is cool with me.

c) If I can see any types of relationships or patterns, I can relay that information to his doctors and let them do what they want with it.

His condition has only been recognized since 1988 so the research is still very young. It's also not life-threatening so it probably doesn't receive as much attention in the academic community or research funding from donors (those are just guesses). Before 1988 it was just diagnosed as fever of unknown origin.

Hopefully, that alleviates some concern? I can tell you that even if you guys helped me create a model that predicted the correct day every month for five years, I would still never give my son the steroids the day before (nor any medication or treatment), solely based on a prediction. To me that seems crazy, but I could see how maybe some people would be tempted to do that. It's really just to crunch some data, brush up on statistics, and try and have a little fun with it. Learn some new things. Maybe it helps, maybe it doesn't. If you met me you'd probably laugh. I'm about the most cautious person around.

Ok, I'm done being long-winded. I've attached his data set if anyone wants to take a look. If you want it in a different format or anything, just let me know. Can't thank you guys enough for taking the time.

Regards,

Michael
 

CE479

New Member
#6
Hi Michael,

Sorry to hear about your son, and well done for looking into this.

Making predictions in medicine is extremely difficult, and people tend to develop prediction scores. They do this by looking at a population, and identifying which variable are associated with the outcome. Then you have to look at which variable are inter-related, and try and come up with a score of independent predictors. How good a predictor a score is depends on its discrimination. I think it will be extremely difficult to come up with such a score for your son, which may then not even be successful, I am sorry to say. What you really need to know is what are the predictors/ trigger factors and you doctor may be able to give advice about this. I would also trust your intuitive knowledge of your son, because he may show certain symptoms/ signs at the start of an episode, so you can give the treatment.

Having done a search of the medical literature, I know that the length of the fever depends on the underlying cause (for example it is shorter in certain subtypes compared to others). Also, temperatures tend to be higher in the evenings. Hopefully the doctors doing the tests will come back to you with more information in this regard. I wish you and your son all the best...