IMPORTANT: This competition uses unique criteria to determine eligible models. It is extremely important that participants read the rules carefully before continuing. In particular, please pay special attention to the Rules section before embarking in this challenge.

Welcome to The NASA Breath Diagnostics Challenge

The objective of this challenge is to develop a classification model that can accurately diagnose patients with COVID-19 based on data that was captured from the exhaled breath of volunteers using the NASA E-Nose device. The total number of patients, and therefore examples, is 63. You will note that this is a limited dataset, so making efficient use of the provided data is of extreme importance in this challenge. We encourage creativity on dealing with the small sample size, since one of the goals of this competition is to address the challenge of limited training data in scenarios such as emerging diseases with few confirmed cases. The limited data also impacts submission testing, so it is very important to understand the rules governing this event and how this will be ultimately scored.

The data consists of 63 txt files representing the 63 patients, numbered 1 to 63.
Each file contains the Patient ID, the COVID-19 Diagnosis Result (POSITIVE or NEGATIVE) and numeric measurements for 64 sensors, D1 to D64. These sensors are installed within the E-Nose device, and they each measure different molecular signals in the breath of the patients.

All sensor data is indexed by a timestamp with the format Min:Sec, which represents the minute of the hour, and the second of that minute in which that sensor was sampled. The hour of the timestamp has been left out, but when the minute counter resets, it means that the next hour has begun. Keep this in mind when working with this time axis.

In order to achieve maximum consistency across patients, the data was exposed to the E-Nose device using a pulsation bag that had previously collected a patient's breath. The E-Nose measurement procedure also includes flushing the sensors with ambient air, which can be used to calibrate the readings taken when exposed to human breath.

The data was exposed to the E-Nose device for all patients using windows of exposure through the following process:

1. 5 min baseline measurement using ambient air
2. 1 min breath sample exposure and
    measurement, using the filled breath bag
3. 2 min sensor “recovery” using ambient air
4. 1 min breath sample exposure and
    measurement, using the filled breath bag
5. 2 min sensor “recovery” using ambient air
6. 1 min breath sample exposure and
    measurement, using the filled breath bag
7. 2 min sensor “recovery” using ambient air
Total time = 14 mins

The data is distributed into training and test sets:
Train: 45 patients
Test: 18 patients

Within the dataset there are also 2 other files: submission_example.csv and train_test_split.csv. The first file represents how all submission files should look like. The values should be 0 for NEGATIVE and 1 for POSITIVE. Failing to follow this format will result in an error or lower score.

The second file (train_test_split.csv) represents which patient IDs are considered for Train (i.e., labeled) and which one are for Test (i.e., not labeled).

The order of the predictions in the submission file should be the same as in the TEST indicated rows in train_test_split_order.csv file.
The index column of this submission file is NOT the ID of the patient, but the order of values in the Result column should follow the one in the train_test_split_order.csv file. This is very important.

The evaluation metric is Accuracy.

The leaderboard will be split into a Public and Private leaderboard, where the preliminary results to advance to the final evaluation stage will be determined by the Private Leaderboard, which will be revealed at the end of the competition period. Again, please refer to the Rules section, especially rules 7 to 11, in order to understand the particular evaluation criteria for this challenge.

Please note that the goal of the Public Leaderboard will be mostly for your reference, as it will represent only an approximate assessment of a model's performance. The final score may deviate substantially from this score. Any attempt to try to "game" the Public Leaderboard score or artificially inflate it will not result in any benefit accounting for the final score and could end in disqualification if the model is found to be purposely overfitting the Public Leaderboard test data.

We wish you good luck in this challenge. If there are questions, please refer to the FAQ, Rules or send an email to [email protected].