Harnessing Real-World Data for Regulatory Use and Applying Innovative Applications

Figures & data

Figure 1 From real-world data to real-world evidence, key capabilities such as health economics outcomes research, real world evidence and population health.

Abbreviations: FDA, US Food and Drug Administration; WHO, World Health Organization.

Table 1 Ten Machine Learning Algorithm Techniques²⁹

Algorithm	Description
Supervised Learning Algorithms	These techniques use training data to learn mapping functions that convert input data to output data. Permits new (non-training) output information to be generated from new inputs.
1. Linear Regression	Determines the potential relationship between input and output variables used for continuous, as opposed to categorical, variables.
2. Logistic Regression	Similar to linear regression but analyzes categorical variables (ie, dichotomous information).
3. Classification and Regression Trees (CART)	Utilizes a decision tree to make predictions for categorical variables, starting with a root node (the first decision), working through internal nodes, and ending with a leaf node.
4. Naïve Bayes	Calculates the probability of an event occurring given that another event has already happened. Bayes’ theorem is utilized. Naïve refers to the assumption that variables are not associated with or independent to each other, which is often not the case in the real world, hence this technique is described as naïve.
5. K-Nearest Neighbors (KNN)	Uses the entire dataset as the training set rather than having training and test data sets. When an outcome for a new data instance is required in this method, the entire data set is evaluated for the closest match to the new data point (nearest neighbor). This is then used to generate new data.
Unsupervised Learning Algorithms	These techniques are used when only input variables are available. There are no output variables. Unlabeled training data are used to model the structure of the results.
6. Apriori algorithm	Generates the rules of association in data sets, eg, data items that frequently occur together in the data set. For example, if a person purchases gift wrapping paper they will often purchase sellotape or sticky tape.
7. K-means	Groups the data into clusters, with centroids calculated for K-clusters. The data points assigned to clusters have the smallest distance between the centroid and all of the other data points.
8. Principal Component Analysis (PCA)	Reduces the number of variables to make the data more straightforward to evaluate. The factor (component) that contributes to greatest degree of variation in the data set is the principal component. The second component captures all of the remaining variance in the data set, which is not correlated with the first component.
Reinforcement Learning Algorithms	These techniques take actions that will maximize reward. Reinforcement algorithms often learn optimal actions through trial and error.
9. Bagging with Random Forests	Constructs multiple models using a bootstrap sampling method in which random subsamples of the original data set are selected to create training data sets. These samples are the same size as the original data, but some values repeat, while other values are excluded. A bagging procedure is then used in which multiple models are created using the same algorithm on the training data sets. The next element is the random forest process in which a random selection of factors is used to find the best factors of where to split.
10. Adaptive Boosting	Builds a series of models independently to correct for the misclassification of data from the prior model.

Figure 2 Benefits achieved with real-world evidence and AI.

Abbreviation: AI, artificial intelligence.

Shaw R. The 10 best machine learning algorithms for data science beginners. Available from: https://www.dataquest.io/blog/top-10-machine-learning-algorithms-for-beginners/. Accessed July 7, 2020.

 Google Scholar