When patients suffer unintended reactions to medicines, it can be both dangerous for the individual and costly to society. However, what if medical professionals could use machine learning to forecast adverse drug reactions (ADRs) and minimise risks to patients?
ADRs are a huge concern within the healthcare industry, directly accounting for approximately 5% of all hospital admissions and estimated to be the sixth leading cause of death worldwide. Personalised medicine is already a fast-developing trend, but imagine if doctors could predict the likelihood of a patient suffering from an ADR before prescribing them drugs.
The FDA's Adverse Event Reporting System (FAERS) collects information on every adverse event recorded worldwide. We used this data to train a machine learning model to forecast ADRs and explore the causes of Omeprazole's adverse drug reactions.
1. Loading the Data into Mind Foundry
The FAERs data set contains a lot of information, but we took extra precautions to exclude variables which presuppose the existence of an ADR, such as the dechallenge/rechallenge codes.
After loading data into Mind Foundry, we are able to see summary statistics of the patients who suffered from ADRs.
In particular, we can see how each ADR - for example hospitalisation - distribute across the other variables.
The factors we shall use to predict the ADRs are the patient's:
- Sex, age and weight
- Dosage amount, form, frequency and route
- Cumulative dose
- Drug manufacturer
The ADRs included in the dataset are:
- Death
- Life threatening condition
- Disability
- Hospitalisation
- Other serious
2. Preparing the Data
Mind Foundry automatically scans the data for potential issues and gives advice to the user on how best to prepare the data for machine learning modelling.
Based on our understanding of the data, we chose to simply mark the missing values in dose amount, frequency and form as "Unknown". We can then automatically apply the advice with a simple click.
Mind Foundry also keeps a full audit trail of every processing step you apply to the data, which can then edit or export to new data.
3. Modelling Adverse Hospitalisations
When we are ready to build our model, all we need to do is tell Mind Foundry which column we wish to predict. It will then automatically handle the rigorous splitting of the data into training and test sets, so it doesn't overfit your solution.
Mind Foundry will then start to search for the optimal data science pipeline and will provide some performance statistics as it does so.
4. Validating and Interpreting the Model
Once it has found the best solution, Mind Foundry validates the health of the model by indicating whether its predictive power is significant or not compared to a random model. These tests are performed on a 10% hold-out from the original dataset which were excluded from the training.
Mind Foundry also indicates the relative weight of each variable in the model's decision making, as well as how each individual variable impacts the outcome.
5. The Findings
Mind Foundry indicates that the chances of a hospitalisation increase with the patient's age, but decrease with their weight.
This might seem strange, but is possible considering that the chances of a disability significantly increase beyond a certain weight and therefore becomes the dominant ADR.
The drug manufacturer also appears to have a slight impact on the chances of a hospitalisation. Finally, the oral route and 40mg dose taken as a tablet will increase the chances of a hospitalisation occurring.
This brief study highlights the potential for machine learning to forecast ADRs using patient data. Looking at the bigger picture, it also reveals the possibility of doctors tailoring prescriptions to the specific conditions of each patient, and drug manufacturers redesigning treatments for at-risk patient populations.
