Hey there!
Have you been:
- working late in the night
- drinking too much coffee
- fighting ADDI’s workspace lag
- spending way too much time debugging stuff
- engineering too many features
- trainining countless boosted decision trees
- continuously crunching log-losses
- And getting super hyped when you finally improve your log-loss by 0.001 ?
(I’ve been doing all that 
)
The Wall 
Sometime, when you face a wall, it’s good to take a step back and think of things differently.
So, I took some time off coding, and went on reading scientific literature on the Clock Drawing Test. 
(Before this AIcrowd challenge, I did not know this test existed, but was immediately drawn to its specifications and ramifications).
There is a lot of information out there… some sources are linked at the end.
Spoiler
reading all this material made me realize how devastating this irreversible disease is…
Just thinking that this disease will slowly destroy your memory and thinking skill is devastating.
Food for thought… You are all bright minds; but 1 of you out of 10 will have Alzheimer over the age of 65. 
Science improving your logloss 
Here are some highlights that I find worthy of sharing and that can have an impact in this competition.
Provide the Circle?
Some clinicians say to give the circle and then ask your loved one to fill in the numbers and hands. There are benefits, however, to providing a blank piece of paper and asking for a clock that shows 10 minutes past 11. If your loved one has to draw the circle, for instance, it is easier to assess “graphical difficulties,” like wavy lines, when more drawing is required. Also, if you don’t use the words “hands” and “numbers,” then more visual memory is required to remember that those are parts of a clock.
Made me think… what if the train data had processed images from more than one source? Could we be analyzing tests that had this circle pre-drawn?
Definitely something to consider when building our features.
The Size of the Clock.
Patients with Parkinson’s disease dementia and Huntington’s disease, a rarer dementia that can develop in a person’s 30s and 40s, are prone to draw a smaller clock (less than two inches), while people with Alzheimer’s disease tend to draw bigger (more than 10 inches) clocks. This is because size awareness is affected differently depending on which illness is present.
Well, there you have it… , size does matter… 
Fixating on Time.
If the person being tested is focused mostly on the time aspect—getting the hands pointed the right way so the clock is correct—to the detriment of the rest of the clock, this is more common in Alzheimer’s disease and Parkinson’s disease dementia than with other dementias.
Graphical Errors.
If the clock looks essentially correct, but is messy (wavy or broken lines), this is more common in Huntington’s disease and vascular dementia than Alzheimer’s disease.
Relates to a previous post from @mmi333,
the label “normal” doesn’t necessarily mean a healthy person, but a person without Alzheimer’s disease
Radical Wrongness.
If the clock doesn’t even look like a clock, like if there are letters or incorrect shapes added, this suggests confusion that is more indicative of Alzheimer’
I wish we had the image processed with an alphabet detection as well 
…
Spatial Errors.
If there are big gaps, numbers outside the circle, or numbers on only one side of the circle, this is more common in Parkinson’s disease with dementia, Lewy body dementia and vascular dementia, and less common in people with Alzheimer’s.
See how it relates to your data?
Did You Know?
As artificial technology advances, it’s possible these evaluations will reveal more and more about the mental state of people taking the CDT.
Right there is this competition objective… improving the future of detection.
Who knows, your own model could one day help detect your own disease? 
Hope you found this helpful.
Have a good week-end.
Cheers!
sources:
How the Clock-Drawing Test Screens for Dementia
The clock drawing-test 1998
Dementia Clock-Drawing Test: Everything Caregivers Need to Know