European Epidemiological Forum

During the presentation the speaker with deal with the following topics:

• Increasingly complex payer landscape for access & reimbursement of new medicines.
• Increasingly complex innovative medicines (CAGT) with curative intent is “breaking” the model.
• Use of AI/ML by payers to integrate RWE and RWE/Real Time Data on conditional reimbursement models.
• Can we afford a cure? Paying for innovation – AI/ML will be the future to integrate value and access.
• How payers are using AI & ML to assess value and deploy innovative reimbursement models.

Risk prediction for Acute Coronary Syndromes (ACSs) represents an unmeet need in medicine. In particular, prediction of risk of death may be of help to tailor aggressive lifestyle interventions and close clinical and with imaging follow up. Moreover, these patients are themselves at increased risk of recurrent thrombotic events, but in the mean time, due to use of antiaggregant therapy they are exposed to elevated risk of bleedings. “Traditional” risk scores, derived with mainly with logistic regression, are derived from patients suffering both ACS and stable angina, consequently with limited reproducibility in every day clinical practice. Moreover, Artificial Intelligence (AI) and Machine Learning (ML), through exploration of non-linear complex relationships may help to capture other kind of associations.

In the PRAISE (PRedicting with Artificial Intelligence riSk aftEr ACS) study, 19826 patients were exploited to predict 1-year mortality, recurrent infarctions and major bleedings. Trials of four ML classifiers [Adaptive Boosting , Naive Bayes, KNN, Random Forest] were employed to generate four models for the prediction of each study outcome. The Adaptive Boosting, which performed superior to others, offered for 1-year all-cause death an AUC of 0.82 (95%CI 0.78-0.85) and 0.92 (95%CI 0.90-0.93). in the evaluation and the validation cohort, respectively. The PRAISE score for MI showed an AUC of 0.74 (95%CI 0.70-0.78) and 0.81 (95%CI 0.76-0.85) in the evaluation and the validation cohort, respectively. The PRAISE score for major bleeding showed an AUC of 0.70 (95%CI 0.66-0.75) in the evaluation cohort and 0.86 (95%CI 0.82-0.89). in the evaluation cohort and 0.86 (95%CI 0.82-0.89). in the validation cohort.

A phenotype definition specifies cohorts of patients that exhibit certain phenotypic traits, such as the same diseases, characteristics, or set of co-morbidities. Electronic health record (EHR)-based phenotype definitions (also referred to as phenotyping algorithms) express the logic required to identify such a cohort from patient health records. Phenotype definitions can be represented in many forms, including narrative descriptions, pseudo-code, or in a standard structured format.

A high-quality definition is one that can be easily implemented across multiple organisations and datasets (portable); one that can be implemented accurately, to facilitate complementary research or validate existing results (reproducible); and one that is shown to effectively capture the disease or condition modelled (valid).

Phenotype libraries provide a platform through which a phenotype definition can be uploaded, stored, indexed, retrieved, and downloaded by users. Rather than just passively cataloguing definitions, these platforms can also impact the quality of the phenotype definitions they host.

One way to do this is for a library to facilitate parts of the phenotype development lifecycle often considered extraneous, including modelling, implementation and validation. For example, libraries that host definitions structured according to some set of standard models, and track the evolution of definitions under these models, often contain phenotypes that are clearer to understand and thus have the potential to be more reproducible.

We shall discuss a number of desiderata for the development of phenotype libraries, to ensure that future libraries are designed to maximise the quality of the phenotypes they contain.