The Covid-19 pandemic has shifted so much attention towards treating the virus that crucial areas of Romanian healthcare have been all but forgotten. Our society presents rampant cronic problems, such as:

1. Child and maternal mortality rates due to poor infrastructure;
2. HIV infestation rates due to the underinvestment into long-time HIV treatment;
3. Teenage pregnancies and the ignorance towards contraception methods due to the lack of sexual health education.

Although the existence of issues such as those listed above is an acknowedged sad reality in Romania multiple succeding Governments have done little to nothing to control and reduce them. Most of the poilticians’ initiatives focus on cutting investments, spendings and salaries from the health and education sectors rather than the allocation of more funds and manpower in order to solve the problems of the people they serve and represent.

We took this opportunity to conduct a visual-focused analysis on the situation of the Romanian heathcare system and its pitfalls, while also drawing a parallel with other E.U. and Schengen countries. We decided to compare our country namely with its Western counterparts, as those are the ones to realize the importance of investing heavily into public healthcare, health education and research. We avoided the United States as it an underdeveloped country when it comes to the healthcare sector.

The presentation contains intuitive visualizations, along with the necesary explanations, conclusions and connections we have made with diferent parts of our data. It also provides a machine learning modelling section in which we showcase how to predict a country’s Human Development Index and how to visually analyze the results. Due to our lack of experience with Rmarkdown, some of the plots have ended up being smaller than expected when embedded in the HTML file. We would like to apologize for the inconvenience and ask the reader to zoom in the browser to check those plots which seem too small.

This project has been developed for the Data Visualization course at the Faculty of Mathematics and Computer Science, University of Bucharest.

The dataset we have employed for this project is the World Development Indicators published by World Bank and forwarded as a CSV file through kaggle. The data set contains 1346 social and economic indicators for 247 countries and geopolitical groups, collected over 55 years, from 1960 to 2015. Thus, it presents a total of 13585 samples.

We have manually processed our data and extracted 79 economic, educational, social and health indicators the we have used within our graphs, along with the country to which they belong to and the year in which they have been recorded. Out of all the preemptively extracted features, we have only used a handful of insighful ones, part of which are presented in the table below:

Bear in mind that the data contains NaN values, so we used imputation methods such as kNN and SoftImpute to infer the missing data points. The first gave an unnormalized output which we used during plotting, whilst the latter gave normalized data, used in the prediction process. Because of the imputation methods, some samples have been skewed. We have taken this into account when visualizing the data and shall notify the reader when required.

We have taken a two-fold approach in our modelling endeavour, by predicting both the GDP per capita and a new composed HDI feature, computed as shown below:

Predicting the GDP was backed up by the fact that it is naturally connected with the wellbeing of the population, by law of casual inference. Our second approach, using the HDI, was of a calculated consistent nature, edorsed by correlations drawn by experts over long periods of time. We have excluded the compound feature that were used to model the HDI feature, namely GNI (current US$), Life expectancy at birth, total (years) and Expenditure on education as % of total government expenditure (%). If we haven’t, then the random forest would obviously use these feature to model the HDI.

We used the random forest model in both cases. The pipelines consisted feature selection using random forest, a K-Fold randomized grid search and finally the prediction itself using the random forest model. We employed the data imputed with the SoftImpute method, as the method normalizes the samples during the process.

The code itself is rather standard for this type of task, so we left it out of the current presentation. However, we did include the optimized hyperparameters for those curious, as well as the top features selected by the model.