Data Science

The promises of Data are great, especially in health and for the healthcare industries: discovery of new therapeutic avenues, acceleration and greater fairness of clinical trials, personalized medicine and optimized patient follow-up, etc.

To help achieve these ambitious goals, Data science needs accessible quality data and above all to interface upstream with the owners of the Data and downstream with the users of the conclusions drawn. Valuing data for companies therefore requires transformations towards a more data-driven model to make the most of the work of Data Scientists.

What opportunities can Data science services generate for my business? Are we equipped and structured today to create this value?

This crucial step in any Data science process determines the success of the analysis at the technical level but also its value at the interpretative level. It may require making choices to increase the signal-to-noise ratio. In particular, the field of Data science focused on the analysis of textual data, Natural Language Processing, or NLP, may require particularly significant data cleaning depending on the source used (example: collecting information on social networks to detect and interpret misspelled words or abbreviations).

What data is rich enough for its analysis to add value? How to extract value from the internal or external databases that we have?

The quality and representativeness of the input data is key to being able to draw relevant conclusions. In particular, poorly balanced or “imbalanced” data can bias learning. If we seek to train an algorithm to classify images of cats and dogs on the basis of 1000 images of cats and 100 of dogs, the notion of a greater frequency of occurrence of cats will have an impact on the classification of new images.

This imbalance can be easy to identify if it concerns the main detection objective but much less so if it is one element among others, for example an over-representation of kittens among the images. Historical databases may be biased, such as clinical trial databases in which Causasian men are over-represented compared to the ratio of the general population. The aim is to identify these biases and correct them by reducing the size of the over-represented sample (undersampling) or by artificially increasing that of the under-represented sample (oversampling).

For example, how can clinical centers be better targeted to achieve good representativeness of the population included?

In the implementation of machine learning models, another technical issue is related to not over-adapting the model to the existing dataset, which is called the phenomenon of overfitting. Indeed, the quality of the model is tested by various indicators that account for the reliability of the prediction such as accuracy (correct detection rates), sensitivity (capacity to correctly detect the “true ones”) and specificity (capacity to detect the “false” ones).

Trying to maximize these indicators can lead to including a lot of variables in the analysis or using increasingly complex models. It is important to keep a sample of your dataset not to train the model but to test it. Knowing that training data is often more homogeneous than real data, it is also important to limit the complexity of the machine learning model chosen to the minimum required. Assembling the results of several models is also a technique to limit the biases inherent in each of the models.

How to adapt a forecast model to anticipate scenarios with events that have never happened in the past?

Finally, the interpretability of model results is a crucial issue. Some of the most powerful machine learning models, deep learning models, do not make it possible to go back to the parameters that allowed the machine to propose, for example, a given classification. It can sometimes be better to have a lower precision of the model but to be able to explain it. For example, if you want to create customer segmentation, it is interesting to know what parameters define the segments in order to then be able to create appropriate interactions and content.

Finally, rigor in the interpretation of the data must be imposed and must be transmitted to the recipients of the results. In particular, it is often tempting to interpret the correlation between two variables as a causal link from one to the other, a conclusion which must most often be based on business knowledge or specific studies in addition to analyses already conducted.

How to make the results of Data science analyses usable and understandable for internal distribution?