OBJECTIVES

: Pharmaceutical consultation (PC) is part of the clinical pathway of all Multiple Sclerosis (MS) patients in our hospital and generates structured data. We propose a new method to systematically analyze PC data and predict relevant outcomes using machine learning (ML) algorithms.

METHODS

: Data of patients between 2016 and 2020 was collected from PC database. The selected features were drug, disease type, gender, age, days of treatment, Expanded Disability Status Scale (EDSS), number of previous relapses and annual relapse rate (ARR). This data was loaded and analyzed using VSCode (Microsoft) on an environment running python 3.8.8 and Jupyter Notebook kernel. Data was analyzed and visualized. The selected outcome to predict was ARR. Data was split in train and test subsamples (ratio 0.3). Models were compared using cross-validation (CV) 10 folds. Selected metrics for model comparison were Mean Absolute Error (MAE), Root Mean Squared Error (RMSE) and R², both for train and test subsamples. The best performing model was selected and hyperparameters were tunned to minimize overfitting, using a GridSearch Method. The final model was then put into production and available through a web-based interface using Streamlit.

RESULTS

: The study included 103 patients, 75.3% were female, mean age of 42 (19-66) and a median EDSS of 1.5. Of all the compared models, XGBoost was the better performing model. To minimize overfitting learning rate was changed to 0.05, max_depth to 3, min_child_weight to 1 and 500 estimators were used. Most important feature was the number of previous relapses (magnitude 0.3). The final model has a MAE of 0.06, a RMSE 0.01 and R² of 0.9.

CONCLUSIONS

: Despite the small dimension of the dataset, it is possible to use machine learning algorithms to predict relevant outcomes with good performance and adequate fit. The use of this model will allow for tailored clinical intervention and better decision support.