PREDICTION MODELS

The interpretation of SQLape indicators is based on the comparison of observed and expected values, taking into account the profile of hospitalized patients. In general, the expected value is the average value calculated among the top half of the Swiss hospitals for the indicator considered. Experience shows that the performance is highly dependent on the risk stratification that is adopted.

Hospitalized patients are grouped into homogeneous risk groups, according to variables that differ from one indicator to the other. For example, it is mainly acute illnesses and operations that explain the length of stay and costs, as well as age, case complexity, and entry mode (transfer from another establishment). On the other hand, to predict the readmission rates, the most relevant variables are chronic pathologies (degeneration, malignant tumors, recurrent diseases), along with age, case complexity, existence of hospitalization in the previous six months, and admission type (scheduled or not). When it comes to predicting the risk of premature death, the specialty involved, combined with case complexity and age are sufficient. As for the risk of complications, it is above all pathology type (degeneration, infection, ischemia, malignant tumor, trauma, and obstetrics) surgical intervention type (organ, invasiveness) that will be decisive, sometimes in combination with age, case complexity and/or length of stay.

The combination of all these variables generates thousands of risk groups, often without a sufficient number of observations to compute precise estimates. Therefore, we aggregate those groups into homogenous strata of risks, in order to obtain a number of observations per group that enables the computation of expected values with no overlapping confidence intervals.

The big advantage of SQLape classification system is that it exploits precise information on all diagnoses and operations in order to design the best prediction model specifically to each indicator. This is not possible with DRG-type indicators, where acute problems are always in the foreground, or Charlson scores, which only include very few co-morbidities. Another advantage is that the risk classes are obtained without over-adjustment (complications and immediate causes of death are not taken into account in the prediction models). Indeed, for example, it is easy to predict the risk of death if it is known that the patient has had a septic shock or a cardiac arrest, but this does not reflect the risk presented by the patient at admission.

Finally, SQLape (R) risk strata are obtained independently of the rank of the diagnosis (main or not). The hierarchy between the diagnostic and operative categories being programmed, it cannot be manipulated by a hospital to improve its results.

These risk strata are regularly updated, according to changes in the reference period (currently 2018-2020). Grouping is carried out intelligibly and follows a medical logic. This makes it possible to obtain strata of risk that are more stable over time.

All predictions are based on statistical control limits, taking into account both observed and expected values random variations, as described in:

Rousson V, Le Pogam MA, Eggli Y. Control limits to identify outlying hospitals based on risk-stratification. Stat Methods Med Res 2018;27(6):1737-1750.

This approach has broader confidence intervals than funnel plots, because it also includes random variation of expected values. However, this loss of precision is somehow compensated by the fact that the random variation due to case mix is not considered here.

More details are given for cost prediction in :
Eggli Y. Introduction à découvrir un nouveau monde (post-DRGs). Chardonne, SQLape s.à.r.l., 2023.

Table. Adjustment variables, specific for each indicator