LOINC Property - Likelihood and associated LOINC terms

List of LOINC terms:

OIMP meeting in Bratislava:

• Likelihood as a new property:

• Likelihood and relationship to Presence

• p=1 → Present, p=0 → Absent

• Likelihood and relationship to dispositions

• p>0 → there is a (pre-)dispostion to X (?)

• Probability of Disease rather or karyotypic morphologic abnormality?

• Angelman syndrome

• Trisomy 21 risk

• 22q11.2 deletion risk

• Modelling:

• | Observable | :

• | Property type | = | Likelihood |, // Likelihood has a precise definition in statistics!

• | Inheres in | = | Fetus |,

• Technique | = | estimation of.. |

• | Component| = | 737542000 | Trisomy 21 (morphologic abnormality) |

• Or

• • • | Towards | = | 41040004 | Complete trisomy 21 syndrome (disorder) | // Not in range
      

Reviewing RELMA, I note that “Likelihood”, “RelRto”, “Probability”,”Risk” are all used in various attributes of LOINC definition of concepts such as we have been discussing.  Reflecting that decision analytic risk/probability instruments are often developed with a specific types of metric as the dependent variable, I suggest that features with technical definitions in probability theory be reserved for specific measurement properties that would allow concise definition of the Observable.  With that in mind, something like the following hierarchy might be reasonable for enhancement of measurement property, given that existing LOINC concepts will often end up with something more specific than ‘Likelihood’ in the definition:

118598001|Measurement Property(qualifier value)|

>>>Expectation risk and/or certainty

>>>>>>Likelihood-Ratio of the odds of occurrence to non-occurrence

>>>>>>Odds-ratio of probabilities of occurrence to non-occurrence

>>>>>>Probability-numeric (scale of 0-1) estimation of certainty of occurrence

I think there are at least 2 different cases we need to consider: 

1. Expression of uncertainty (or certainty) of an observation result. The 75668-4 example seems to contain two distinct observations: one of the type of sex of the fetus by the "Dosage of chromosome specific cf DNA" and then one report of the certainty (or lack of uncertainty) of the result of the former observation. There should be a generic way of expressing uncertainty of observation, likely one which involves the information model, e.g. as in ISO 21090 QTY and UVP data types (or in FHIR).
2. Prediction or prognosis of a possible future event. There would be a spectrum of the amount and kind of data this prediction is based upon. A prediction based on solely on population data (without any patient-tailored analysis beyond classification into the right population) would be the same as the prevalence of the event. Other case where more patien data is used for making the prediction include prediction instruments such as CHA₂DS₂-VAScscore for stroke risk for AF patients or cardiac surgery outcome prediction scores.

The former (1) could be represented similar to a presence observable with PROPERTY like Jim's suggestion and use of INHERES IN and COMPONENT attributes (if the information model representation is not preferred). The latter (2) could be represented similar to function (realizable) observables, possibly with a Prevalence property (to be added to Jim's list) if applicable, with INHERES IN and HAS REALIZATION (which then needs range extension) attributes.

There are currently 60 'Risk' type components with Property=LP185777-2 Likelihood in LOINC v2.63. As Farzaneh's earlier post shows, they're mostly related to the risk of various chromosomal disorders.

But there are a roughly equal number of LOINCs that look to be concerned with clinically very similar Scale_type=Qn notions of 'risk', but where the property isn't given as 'Likelihood'.

For example these from the cardiovascular risk stable:

Perhaps these are content errors in current LOINC?

If not, then there may be more than discrete notion in LOINC of quantifiable 'risk' and likelihood, and so we may need to clarify what those are before trying to represent one or the other in SNOMED.