I’m working with segmented cells from thin blood smear images and using deep learning models to classify parasitic and uninfected cells. I obtained the following values: Accuracy: 0.986, AUC: 0.99, Sensitivity: 0.981, Specificity: 0.992, F1-score: 0.987, MCC: 0.972, TP: 1329 , FN: 26, FP: 11, TN: 1364. In one of my submissions, I have got a question from a reviewer like this:
Given a system (with p sensitivity, r specificity), how many RBCs need to be analyzed to confidently decide that a person is infected by malaria (given he has q% parasitemia). With the above reasoning/study, report how many RBCs you need to analyze to confidently decide (statistically significant) a person (with q% parasitemia) is infected or not by the best system that you have developed. Also report by the same system, what is the lower level of parasitemia you can diagnose.
Can anyone suggest me an answer (assuming the degree of parasitemia) since we don't have that measure.
A simple way to knowing that, from sensitivity specificity perspective,
You could add "the 4 input value" to epiR class in R, and look at resultant number to diagnose (NTD), AKA minimum repeated diagnosis to confidently says that a blood sample belong to a group or another. :)
Let say, for example, NTD is 2.5, then You need 3 measurement repeats, and, depending to Your methods, agregate the result as average or most likely, then multiplied by RBC needed per diagnosis, which, in Your case, I assume one (Correct me if I am wrong, please). :)
https://rdrr.io/cran/epiR/man/epi.tests.html
@Jenie: I apologize I couldn’t understand. Could u please give an example with the performance metrics measurements I have given? My question is how many slides we have to analyze in the field to get a reasonable measure given the sensitivity, specificity values and degree of parasitemia.
@Rajaraman,
There one problem, I don't know how Your solution works, and the exact sensitivity specificity input values that You have used.
I answer anyway. :)
Given Sensitivity: 0.981 and Specificity: 0.992, I assume You have tried to measure 1000 high parasitic cell samples and catch 981 of them, and You measure 1000 normal or low parasitic cell samples and catch 992 of them. Recalculate using real value later, please. :)
With confidence level of 0.95, that would yields results,
sensitivity 9.810000e-01 9.704876e-01 9.885230e-01
specificity 9.920000e-01 9.842980e-01 9.965400e-01
number to diagnose 1.027749e+00 1.015164e+00 1.047356e+00
That would yields that, You need 1.03 times of measurement repetition. Rounding up, that means, You should measure 2 times per diagnose. :)
Multiply the result with how many RBC per diagnosis needed, then You have got Your answer. :)
Fell free to contact me if You have anything to ask. We all comrade here. :)
- Badges
- Science topic