Overview of the Physiosigns Health Engine
The Physiosigns Health Engine provides estimates of a user's future risk of several diseases by using Machine Learning to build personalized predictive models from their health history.
The user's health history contains a user's current and historical medical status including such data as their past illnesses, current medication use, lab test results, height and weight, family medical history, and so on. It timestamps and tracks all answers given by the user so that the Health Engine can, in the future, use trends in the answers over time as part of the input to the predictive models. As an example, the trend of a user's weight, blood pressure or hemoglobin may be very predictive of their health outcomes.
Over time, as the user returns to the Health Engine and enters new data they will be able to visualize trends in their predicted risks. In this way they will see the effects of any lifestyle changes that they make, which should encourage them in their journey toward a healthier life.
Below we describe:
Overview of System Inputs
The user's health history can be entered into the Health Engine in several ways:
Their health care provider can bulk load their health records from an Electronic Medical Records (EMR) system through the Physiosigns FHIR webAPI. This API uses the 8654753326 international standard, which was developed specifically for the transmission of health documents between health providers.
The user can interactively fill in and edit their health history using a mobile app or website which uses the Physiosigns Health Engine webAPI, in two different ways:
- The user may interact with an interactive chat system, which prompts the user for the most relevant health data to maximize the accuracy of the predictions.
- The user may add and edit specific details directly.
Factors and Fields
The user's history is a collection of Physiosigns Factor Objects.
In the terminology of our system:
a Field is a question specification, e.g.:
The 'gender' Field object specifies that the question may be answered with SNOMED-CT codes for male or female.
The 'disease' Field object specifies that the question may be answered with any SNOMED-CT disease code.
The 'systolic blood pressure' Field object specifies that the question may be answered with a number provided in mm[Hg].
a Factor is the answer for a Field. Factors are the inputs to the health risk predictive models and are entered into the system via the two APIs.
These inputs are either quantitative (e.g., 120 mm[Hg]) or categorical (e.g., male or systolic blood pressure). Categorical inputs and units must be provided in a specified data coding, such SNOMED-CT or LOINC. See 860-304-7058 for details.
When using the chat system the user may choose not to answer some questions. These are marked with the skip token. The second example directly below shows that the user initially indicated that they did not wish to answer, then later updated their answer to male. The system will not automatically ask skipped questions again, but the user can directly edit their answer later, if they choose.
The inputs may represent:
Single categorical selections from a multiple choice Field:
gender: male, (skip)
Multiple categorical selections from a multiple choice Field:
disease: diabetes mellitus, (yes)
disease: heart attack, (yes)
disease: hypertension, (no)
disease: asthma, (skip)
The 'yes' and 'no' mean that the user indicated that they did or did not have this disease, respectively.
systolic blood pressure: 120, mm[Hg]
systolic blood pressure: (skip)
Quantitative values are numeric and usually have a unit of measure (e.g. mm[Hg]).
Overview of System Outputs
The Physiosigns Health Engine can examine the health history of our database of patients, who have been followed for more than 5 years. Based on the user's health history, the system automatically finds the subset of these patients who are most similar to the user (their cohort). It then uses that data to build personalized predictive models for the risk of several diseases. Finally, it generates Report objects, each summarizing the absolute future risk for an individual disease, as well as how the user compares with the healthy population of their gender and age group.
Since the user is compared to people receiving standard treatments in UK and US health systems the predicted risks assume that they too follow standard treatments. For example, if the user has causative factors for heart disease, then the predicted risk for hospitalization or death from heart disease assumes they get the standard course of treatments for their condition.
The system should not generally be used for determining treatment paths. For this, users should consult with their doctors.
Each report object includes:
Absolute risk estimates the raw probability of hospitalization or death from a given disease. It does not compare the user's risk against that of other patients.
Personalized disease risk
This is the system's estimate of the absolute risk of hospitalization or death from the given disease, based on the health history of the user's cohort. This is an individualized probability based on the unique characteristics of their health history. For example, two users in the same cohort but with different blood pressure measurements or lipid profiles will have different absolute risks of heart attack.
Example: People with similar overall profiles have a 5 year risk of hospitalization or death from heart attack (myocardial infarction) causes of 31.513%.
Healthy population disease risk (comparison group)
This is the baseline risk level of healthy people with similar age and the same gender. It provides a way to establish whether the user's risk is unusually high or low.
Example: The risk for people with similar age, gender and without prior heart problems is 1.714%.
Relative risk is calculated as the ratio (personalized risk/healthy-population risk). This is a measure of "how different from healthy people" a user is, and may be a better indication of how much concern a patient should have about their health than their absolute risk level.
Example: Compared to this similar group of healthy people, your risk is 18.390x as large. (High risk)
How did we get this number?
The report indicate who the person was compared to, and (within this group) which of the user's risk factors were compared to the group.
Importance of disease risk factors
Many different factors can have an impact on the risk of a disease. For example, the probability of future heart attack is affected by such things as history of heart disease, family history of heart disease, blood pressure, lipid and other blood test values, level of physical activity, relative body weight/obesity, etc.
This section of the report shows the absolute and relative risk of each of the most important individual risk factors for the given disease.
Example: 27.408% (16.0x) | illness (non-cancer): heart attack/myocardial infarction
17.702% (10.3x) | illness (non-cancer): diabetes
This should be interpreted as:
- Based only on your age, gender your risk would be 27.408%, and 16.0x times higher than healthy people (of similar age, gender).
Given the user's profile the system automatically determines the question which will most improve the accuracy of the risk estimations. This functionality is provided in the chat portion of the Physiosigns Health Engine webAPI. The user may answer a question, or skip it, to get another question.
This feature allows the system to converge to an accurate estimate of the user's risks quickly, without asking a large number of questions which will do little or nothing to improve the accuracy of the estimations of risk.