Predicting health outcomes using data modeling approaches is an emerging field that can reveal important insights into disproportionate spending patterns. This book presents data driven methods, especially machine learning, for understanding and approaching the high utilizers problem, using the example of a large public insurance program.Five years ago I explained in this blog our experience on predictive modeling. This a key reference book.
If the top 5% patients at risk of becoming HCUs are followed, the achieved sensitivity and specificity is 42.2% and 97%, respectively. These values suggest very reasonable predictive power, indicating that the model picks up 42.2% of all high-cost healthcare users and correctly identifies 97% of those who are not high users.Catalonia results:
En el modelo, todas las variables fueron estadísticamente significativas excepto el sexo. Se obtuvo una sensibilidad del 48,4% (intervalo de confianza [IC]: 46,9%-49,8%), una especificidad del 97,2% (IC: 97,0%-97,3%), un VPP del 46,5% (IC: 45,0%-47,9%) y un AUC de 0,897 (IC: 0,892-0,902).The models are slightly different, while the results are close.
Methods: We estimated a new SLT/L-GoM model of the natural history of AD over 10 years in the Predictors 2 Study cohort: N = 229 with 6 fixed and 73 time-varying covariates over 21 examinations covering 11 measurement domains including cognitive, functional, behavioral, psychiatric, and other symptoms/signs. Total remaining life
expectancy was censored at 10 years. Disability was defined as need for full-time care (FTC), the outcome most strongly associated with AD progression. All parameters were estimated via weighted maximum likelihood using data-dependent weights designed to ensure that the estimates of the prognostic subtypes were of high quality.
Goodness of fit was tested/confirmed for survival and FTC disability for five relatively homogeneous subgroups defined to cover the range of patient outcomes over the 21 examinations.
Results: The substantial heterogeneity in initial patient presentation and AD progression was captured using three clinically meaningful prognostic subtypes and one terminal subtype exhibiting highly differentiated symptom severity on 7 of the 11 measurement domains. Comparisons of the observed and estimated survival and FTC disability probabilities demonstrated that the estimates were accurate for all five subgroups, supporting their use in AD life expectancy calculations. Mean 10-year TLE differed widely across subgroups: range 3.6–8.0 years, average 6.1 years. Mean 10-year DFLE differed relatively even more widely across subgroups: range 1.2–6.5 years, average 4.0 years. Mean 10-year DLE was relatively much closer: range 1.5–2.3 years, average 2.1 years.Excellent, good job from Duke University, where I did part of my PhD, using the same methodology Grade of Membership.
These common interpretations of end-of-life spending flirt with a statistical fallacy: Those who endup dying are not the same as those who were sure to die. Ex post, spending could appear concentrated on the dead, simply because we spend more on sicker individuals who have higher mortality—even if we never spent money on those certain to die within the year. Empirically, this suggests using predicted mortality, rather than ex post mortality, to assess end of-life spending.
Less than 5% of spending is accounted for by individuals with predicted mortality above 50%. The simple fact that we spend more on the sick—both on those who recover and those who die—accounts for 30 to 50% of the concentration of spending on the dead.Crucial conclusion:
In sum, although spending on the ex post dead is very high, we find there are only a few individuals for whom, ex ante, death is near certain. Moreover, a substantial component of the concentration of spending at the end of life is mechanically driven by the fact that those who end up dying are sicker, and spending, naturally, is higher for sicker individuals. Of course, we do not— and cannot—rule out individual cases where treatment is performed on an individal for whom death is near certain. But our findings indicate that such individuals are not a meaningful share of decedents. These findings suggest that a focus on end-of life spending is not, by itself, a useful way to identify wasteful spending. Instead, researchers must focus on quality of care for very sick patients.Good article.
Use of Latent Class Analysis and k-Means Clustering to Identify Complex Patient Profiles
Instead of predictive modeling using costs, this is the right approach from a clinical point of view:
This cohort study analyzed the most medically complex patients within Kaiser Permanente Northern California, a large integrated health care delivery system, based on comorbidity score, prior emergency department admissions, and predicted likelihood of hospitalization, from July 18, 2018, to July 15, 2019. From a starting point of over 5000 clinical variables, we used both clinical judgment and analytic methods to reduce to the 97 most informative covariates. Patients were then grouped using 2 methods (latent class analysis, generalized low-rank models, with k-means clustering). Results were interpreted by a panel of clinical stakeholders to define clinically meaningful patient profiles.
And the figures below reflect these results.
Great article.
Figure 1. Seven Patient Profiles Derived From Latent Class Analysis
Figure 2. Comparison of k-Means Clustering With Latent Class Analysis (LCA)
Table 1. Baseline and 1-Year Follow-up Characteristics of the Overall Population and by Patient Profile
