A small proof-of-concept machine learning project that predicts the yearly admission cutoff score at the University of Khartoum (UofK) for Medicine and Dentistry, using the average score of the top 100 students in the Sudanese Secondary Certificate.
This is a small experiment, not a production model. It was built on a very small, self-collected dataset to test whether the idea can work. The results suggest it is promising and would benefit from more data.
Every year, families and students mostly rely on the previous year's admission results to decide which programs to apply for. But cutoffs can change a lot from one year to the next, as happened in 2020, and many students were surprised by their results.
The long-term goal of this project is to give students a data-based estimate of the coming cutoff, so they can choose programs they have a realistic chance of being accepted into. The current version is only a first test of whether this is possible.
The project covers two programs at UofK:
- Medicine
- Dentistry
I chose them because they usually require the highest scores for admission at UofK, so the performance of the top students in the country is a relevant signal for predicting their cutoffs.
No public dataset exists for this problem, so I collected the data myself, which was difficult. The data is yearly: one row per year, with the national exam statistics, the average score of the top 100 students, and the UofK cutoff for each program.
- Feature used: the average of the top 100 scores each year. More features could be derived from the top-100 scores, but this project uses only the average.
- Target: the UofK admission cutoff for Medicine and for Dentistry.
- Training years: 2016-2022 (7 years).
- Excluded years: 2023-2024, because the number of accepted students dropped by about half due to the war in Sudan.
- Test year: 2025.
The dataset is very small and is not publicly available, so it is not included in this repository.
For each program, a simple linear regression was trained on the top-100 average. The feature was chosen because of its high correlation with the cutoff in the training years:
| Program | Correlation with top-100 average |
|---|---|
| Medicine | 0.92 |
| Dentistry | 0.86 |
Predictions for 2025, compared with the actual cutoffs. The baseline simply predicts the average cutoff of the training years, to check that the model adds value beyond that.
| Program | Training years | Predicted | Actual | Model error | Baseline error |
|---|---|---|---|---|---|
| Medicine | all 2016-2022 | 94.78 | 94.7 | 0.08 | 0.86 |
| Medicine | without 2020 | 94.48 | 94.7 | 0.22 | 1.17 |
| Dentistry | all 2016-2022 | 93.23 | 92.3 | 0.93 | 0.06 |
| Dentistry | without 2020 | 92.43 | 92.3 | 0.13 | 0.45 |
Medicine was predicted closely and beats the baseline with or without 2020.
Dentistry is more sensitive to the training data. With all years included, the model is less accurate than the baseline. 2020 is an outlier year (the highest top-100 average in the data, 97.0, together with unusually high cutoffs), and a single year like this can strongly change the fitted line. Without 2020 the model predicts 92.43 against an actual 92.3 and beats the baseline.
Both versions are reported so the effect of that one year is visible.
- Only 7 training years and a single test year (2025), so these results show that the approach is feasible, not that it is reliable.
- A single unusual year, such as 2020, can noticeably change the results.
- Only one feature (the top-100 average) is used.
- Admission cutoffs also depend on factors that are not in the data, such as seat capacity, quotas, and policy changes.
- The dataset is not public, so the notebook cannot be re-run from this repository alone.
- Collect more years and more data
- With permission from the relevant ministry, collect the planned number of seats per program, since seat capacity is likely an important factor in the cutoff that is missing from the current data
- With the same permission, use the scores of all students (anonymized) instead of only the top 100, which would allow richer features such as the full score distribution
- Build more features from the scores (for example the median, minimum, and spread)
- Predict a range with an uncertainty estimate instead of a single number, so students can judge how risky a choice is
- Add other programs and universities
- Use leave-one-year-out validation and compare with simple baselines
Python, Pandas, Scikit-learn, Matplotlib, Seaborn, Google Colab