Notes from a lecture conference by Tao ZHANG of Shandong University, introduced the types of clinical prediction models, the classification of clinical prediction models research and the research process of clinical prediction models.

Clinical prediction model

Types of clinical predictive models

Clinical predictive modeling refers to the use of multi-factor models to estimate the probability of having a disease or the probability of a future outcome.

Diagnostic model

If clinical signs and symptoms are present, consider the diagnostic model. The predictors include clinical symptoms and signs, imaging examination, laboratory examination, etc., it’s a horizontal prediction, and the outcome is whether or not there is a disease.

Prognostic model

If a disease is identified, consider the prognostic model. Predictors include patient characteristics, disease characteristics, imaging examination, laboratory examination, etc., it is a longitudinal prediction, and the outcomes include death, recurrence, disability, complications, etc.

Tertiary disease prevention system

• Primary prevention of disease. Clinical prediction models can provide patients and doctors with quantitative risk values for future diseases based on current health status, and provide more intuitive and powerful scientific tools for health education and behavioral intervention. For example, the Framingham Cardiovascular Risk Score, based on the Framingham Heart Study, has established that lowering blood lipids and blood pressure can prevent myocardial infarction.

• Secondary prevention of disease. Clinical prediction models, especially diagnostic models, can often provide diagnostic schemes with high sensitivity and specificity with the help of non-invasive, low-cost and easily collected indicators, so as to practice the disease prevention concept of “early onset, early diagnosis, and early treatment”, which is of great health economics significance.

• Tertiary prevention of disease. The prognostic model can provide quantitative estimates of the probability of disease recurrence, death, disability, and complications, so as to guide the formulation of symptomatic treatment and rehabilitation programs, prevent disability and promote functional recovery, improve the quality of life, prolong life, and reduce mortality.

Stage of disease

Prevention:

• Identification of high risk factors
• Identification of high-risk groups

Diagnosis:

• Symptoms
• Physical signs
• Physiological and biochemical indicators

Treatment:

• Need for Treatment?
• The choice of treatment

Prognosis:

• What are the outcomes after treatment
• Which factors influence outcomes

Clinical prediction model (CPM)

• Diagnostic model
• Prognostic models

Randomized controlled clinical trial (RCT)

Classification of clinical predictive model studies

• Studies of searching for diagnostic / prognostic factors
• Prediction model development without external validation (internal validation)
• Prediction model development with external validation
• Validation study of the prediction model
• Impact studies of prediction models

Research process of clinical predictive model

1. Preparatory work
   • Establish research question
   • Select data source
   • Data preprocessing
2. Statistical analysis
   • Model selection
   • Single factor analysis and transformation of predictor variables
   • Multifactor analysis and screening of predictor variables
   • Fit model / calibration model
   • Evaluating model performance (Model validation)
3. Results display
   • Presentation prediction model
   • Report the findings

Research design

1. Diagnostic model
   • Study design type: cross-sectional study / cross-sectional data of cohort study
   • Study population: a population with certain symptoms and signs that are suspected of having a disease
   • Diagnostic factors: symptoms, signs, examination test results
   • Outcome: presence or absence of current illness
2. Prognostic model
   • Study design type: cohort study / cohort in RCT
   • Study population: no current disease, chance of future disease / current disease, what will the future prognosis be
   • Prognostic factors: socio-demographic characteristics, disease history, medication history, physical examination results, imaging, electrophysiology, blood and urine sample examination, pathological examination, disease stage and characteristics, and omics
   • Outcome: future illness, death, recurrence, disability, complications, etc
3. Sample size - Limitations of small sample studies
   • The stability and reliability of the study are poor
   • Missing values have great influence on parameter estimation in small sample studies
   • Stepwise screening of predictors is common, but risky in small sample studies
   • Lack of effective validation, especially using split data as validation set

Data preprocessing

1. Outlier processing
   • Consider truncation at 1% and 99% subsites
     • If the subsites are less than 1%, reassign the value to the value of 1% subsites
     • If the subsites are greater than 99%, reassign the value to 99% subsites
   • For data that is clearly not normally distributed, variable transformations should be considered before outliers are checked
   • High-dimensional data can be reduced to determine outliers
2. Missing value processing
   • Several different deletion mechanisms
     • Missing Completely at Random (MCAR)
     • Missing at Random (MAR)
     • Not Missing at Random (NMAR)
   • Missing value handling method
     • Complete case study: Simple and straightforward, but it loses a lot of samples
     • Univariate interpolation: It is convenient and fast, but increases the concentration of variables and affects model fitting
     • (model-based) single interpolation: The correlation between variables is used, but the single interpolation model has certain randomness
     • (model-based) multiple interpolation: Randomness can be measured, but the subsequent modeling methods are complex

Statistical analysis

Choose the right predictors

Model evaluation index

Model comparison index

Results display