Clinical Decision support systems involve any system that supports the clinical field. When assessing a patient, there is a lot of data to go through in the electronic health records, and doctors cannot process and go through all the data. Computers however can process this information quickly, and can notice small changes in the data as well. Clinical decision support systems are used in diagnosing patients, offering treatment plans, predicting a time of death, readmission rate, transplant survival and more. To achieve this, mathematical models, logistic regression, deep learning and larger data sets are used.
There are many different types of clinical decision support systems. There are many different ways that the systems are categorized. They are categorized by system function, advice giving, style of communicating, interaction with others and the decision making process. The system refers to whether or not the system classifies something, such as a diagnoses, or if it offers something to do, like a treatment plan. The style of giving advice refers to whether the clinical decision support system uses passive advice or active advice. Passive advice is advice that is give through some human interaction. For example, by clicking a tab or buttons. The style of communication refers to whether or not the system consults or critiques. Consulting is where the computer gives advices to the patient on what choices they should make. For example, the patient will tell the computer a diagnoses they have received, and the computer will offer a treatment plan. Critiquing is when the patient makes choices and the computer will alert the patient if these choices aren’t good. The way the system interacts with the user refers to whether or not the user clicks tabs, or buttons, or if there's voice recognition. It refers to any way that the system interactions with the user. The decision making process refers to the way that the systems make their predictions. This could range from simple flow charts to more complex artificial neural networks, Bayesian, support vector machines or AI.
A challenging part of this is that many people have trouble trusting these model's predictions. Additionally, there isn't a lot of widespread use of clinical decision support systems. This is because more simple models don't look at multiple patient characteristics, and even more advanced models don't give alerts specific to a certain patient, rather all the alerts specific to a certain treatment. Doctors also override 49%-96% of alerts, because the alerts often disrupt the work and aren't very specific.
To try and combat these issues validation is a very important step in this process. There are 4 different types of validation. Technical validation, therapeutic retrospective validation, pre implementation prospective validation and post implementation prospective validation. Technical validation involves making sure that the predictions of the model are correct. Therapeutic validation involve making sure that the predictions are helpful and relevant to the clinical field. Pre implementation prospective validation involves implementing the model in practice to see if it is effective. Post implementation prospective validation involves maintenance of the system and making sure to improve it overtime.
For this week, I used K Nearest Neighbors on a dataset about tumors to make a model that predicts which tumors are benign and which are malignant.
Kaggle Dataset: https://www.kaggle.com/datasets/adhamelkomy/breast-cancer
Tumor Prediction Code