In traditional studies, buyers can reflect mechanism with relation to selection of participants, the variables the outcome and the analysis. This buyer should be carefully assessed and identified. Nevertheless, electronic health records have inherently biased due to the fact that sick people, they tend to be monitored more often. This later type of buyers is difficult to be interpreted, although it might carry useful information. Prediction models in clinical applications are normally either diagnostic or prognostic. In diagnostic models we look in the probability of a certain condition being present but not yet detected, since for these models, the prediction is for an outcome already present. The prefer design is a cross sectional study. There are several examples of diagnostic models, for example, cancer detection or L diagnosis of alzheimer's. On the other hand, prognostic prediction models estimate whether a patient or an individual will experience a specific event or outcome in the future. Normally, this is within a certain time period that can range from minutes to hours, days, weeks, months or even years. So the relationship here is longitudinal. Prognostic models are not only used for patients, but they're used also for healthy individuals to identify the risk for a disease, for example, for pregnant women. It is interesting to be able to predict the risk for diabetes. It is not surprising that both diagnostic and prognostic models are useful in a number of fields and they are used in cancer neurology and cardiovascular disease extensively. Despite the different timing of the predicted outcome, diagnostic and prognostic prediction models have many similarities. For example, the type of outcome is often binary. In a diagnostic model, this will translate, for example to a target condition being present or not, whereas in a prognostic model will translate whether an outcome event will a cure in the future or not. Furthermore, in both diagnostic and prognostic models, the key interest is to estimate the probability of an outcome being present or a curing in the future. And this is based on multiple predictors with with the purpose of informing patients and individuals in quieting decision making. We've discussed that predictive models can be evaluated based on their discrimination and their calibration. Nevertheless bias can be present and create systematic errors In a study. Bias may reflect shortcomings in the study design or in the way the study took place or in the analysis of the data. In order to identify bias, we need to consider all these aspects. Therefore we need to think about the intended use of the model, the participants that it targets, the predictors used in the modeling ,as well as the predicted outcome. Taking all these considerations together, it has been suggested that assessing buyers and clinical predictive models should involve four steps. This involves assessing bias in the selection of participants, the selection of variables and predictors, the outcome in the analysis. Additional appraisal should consider also there is a question, for example whether the included participants or settings much the question at hand. Similarly, the definition assessment, and timing should be relevant to the research question. In a recent study, it has been argued that a tool, in fact a set of questions is important to assess bias in clinical predictive models. This questioner includes four groups of questions. The first one is related to the participants, are the data consider a ticket to answer various question. And we're all inclusion and exclusion criteria for participants appropriate, an example of risk of bias here would be to remove patients due to missing values. If the missingness in the data is not random, then this will create bias in the data collected. The set of consideration in terms of the predictors involved on whether the predictors are defined and assessed in a similar way across all participants and whether they are present at the time the model is intended to be used. It also involves the fact that there may be biased in the research team analyzing the predictor assignment. Therefore, predictor assessment should be made without knowledge of the outcome of the data. In the outcome category we see a set of questions. These questions are important to determine whether the outcome has been defined appropriately and also whether it was a prespecified or standard outcome definition used. Importantly, the outcome should be defined and determined in a similar way for all participants. In normal predictors shouldn't be excluded from the outcome definition. Finally, the time in terrible between predictor assessment and outcome determination should be appropriate in order for a study to be meaningful. For the analysis it is important to look on whether the sample size is articulate and how continues and categorical predictors has been handled. Here Ii's also highlighted that Selection of predictors based on uni-variate analysis should be avoided. Questions here also involved performance evaluation as well as assessment of over-fitting. The authors recommend rating the prediction model as having low risk and buyers assessment. If no relevant shortcomings were identified based on these questions. If at least one domain have been identified with high risk of buyers, then a novel judgement of high risk of buyers should be used. Similarly, medium risk of buyers should be assigned if at least one of these factors have medium risk. If a prediction model is developed without any external validation on different participants, then it should be considered to downgrade it to high risk of buyers even if all four domains had been identified as law. The exception to this rule is if the model development is based on a very large data set. Electronic health records are collected in an irregular fashion and it's natural that they carry more information about people that they are sick. If there are more electronic health records for a patient, this is not at random but it reflects that the subject is seriously ill. The presence or absence of patient data is potentially informative with respect to the individual health condition. This is called informative presence. So the presence or absence of a patient's data at any given time point carries information about their health status. In other words, we can build a model that predicts outcome just based on whether a diagnosis took place or not. Informative presence is fundamentally different than missing data, since there is no intention of collecting the data from people that they are healthy. Informative observation in electronic health records referred to the timing frequency and rate of patient longitudinal pattern of observations that carry information about how their health state evolves. In other words, informative observation can be seen as the continuous process that generate informative presence. Some work has focused on this phenomenon and it has seen it as a problem because of the statistical challenges it imposes in the estimation of a causal or association studies. This is an example of how informative presence can impact prediction. On top we see the systolic blood pressure of a patient and we see both observed as well as un observed values. We see that the only value that is observed is relatively high. On the other hand, patient two has no observed values and this possibly reflect that there was no clinically need to take a blood pressure measurement at any time because his historic blood pressure is within normal range. On the right we see another example of informative presence based on data from the mimic database. One of the patients has far more blood glucose test during his admission to the intensive care compared to the other one, and this is likely due to the fact that his blood glucose is much higher and much more variable. In other words, a more severe condition often results in more intense monitoring. There is a view that informative presence and informative observation are opportunities to draw information from the electronic health record that is not explicitly recorded. Summarizing, it's of paramount importance to identify risks of buyers in order to assess clinical predictive models. Nevertheless, electronic health records include inherently informative presence and informative observation bias. This makes interpretation of the results far more challenging. Nevertheless, research status have used this as a source of information to improve prediction.