Skip to main content

Table 4 Predictive Ratio and Net Compensation Values of Prospective Machine Learning Models on SDH-Based Subgroups in the Test Set

From: Incorporating machine learning and social determinants of health indicators into prospective risk adjustment for health plan payments

   

Model Predictive Ratiob and Net Compensationc

Subgroup

No. (%)

2017 Spending ($)a

ML (95% CI)

ML with SDH (95% CI)

Total

117,616 (100)

6677

1.000 (0.976, 1.024)

1.000 (0.976, 1.024)

0 (− 105, 105)

0 (− 105, 105)

Poverty

 Median Income in the Past 12 Months, $

4923 (4.2)

10,818

1.017 (0.915, 1.120)

1.006 (0.905, 1.108)

− 183 (− 836, 470)

−67 (− 729, 595)

 Families Under 0.5 Ratio of Income to Poverty Level in the Past 12 Months, %

7932 (6.7)

9344

0.966 (0.882, 1.050)

0.948 (0.865, 1.031)

331 (− 138, 801)

510 (33, 987)

 Families Between 0.5 and 0.74 Ratio of Income to Poverty Level in the Past 12 Months, %

6651 (5.7)

8952

1.010 (0.912, 1.108)

0.988 (0.892, 1.084)

−89 (− 599, 420)

109 (− 408, 627)

 Families Between 0.75 and 0.99 Ratio of Income to Poverty Level in the Past 12 Months, %

7194 (6.1)

9395

1.052 (0.956, 1.148)

1.010 (0.919, 1.101)

− 467 (− 977, 43)

−94 (− 613, 425)

 Families Received Food Stamps/Snap in the Past 12 months, %

9009 (7.7)

9001

1.028 (0.941, 1.115)

0.996 (0.912, 1.079)

− 247 (− 684, 191)

39 (− 409, 487)

 Population Unemployed, %

10,278 (8.7)

7055

0.961 (0.886, 1.036)

0.957 (0.882, 1.032)

289 (−71, 649)

316 (−51, 683)

 Gini Index of Income Inequality

16,155 (13.7)

6138

1.054 (0.985, 1.122)

1.021 (0.955, 1.087)

− 312 (− 578, −46)

− 126 (− 393, 140)

Education

 Population Obtained High School Diploma, %

9482 (8.1)

7555

0.987 (0.900, 1.073)

0.974 (0.889, 1.058)

102 (− 324, 529)

205 (− 227, 637)

 Population Obtained Bachelor’s Degree, %

4169 (3.5)

11,338

1.032 (0.923, 1.142)

1.027 (0.917, 1.136)

−353 (− 1139, 433)

− 294 (− 1080, 492)

Other

 Population Speak English Less than “Very Well”, %

23,659 (20.1)

5453

1.023 (0.963, 1.083)

0.989 (0.932, 1.046)

− 124 (− 346, 98)

61 (−161, 283)

 Families with Single Parent, %

9097 (7.7)

9880

0.993 (0.910, 1.076)

0.978 (0.896, 1.060)

65 (− 397, 527)

224 (− 246, 693)

 Population Without Health Insurance Coverage, %

13,656 (11.6)

8333

1.066 (0.990, 1.142)

0.990 (0.921, 1.059)

− 516 (− 885, − 147)

83 (− 287, 454)

  1. Comparison of machine learning prospective risk adjustment models without and with the addition of SDH indicators as predictors (see Table 1 for a complete list of SDH indicators). The predictions for each model were adjusted so that the mean of the predictions over the total test population was equal to the mean of the actual costs, resulting in a predictive ratio of exactly 1.0 over the total test set population. Subgroups were composed of members in the lowest decile of ZIP codes with respect to the corresponding SDH variable (see Supplementary Information Table S1). Only socioeconomic variables are considered in this subgroup analysis, and results on age and sex subgroups are shown in the Supplementary Information
  2. aSpending included all healthcare utilization in 2017 of members with full enrollment in 2016 and 2017. Values larger than $400,000 were replaced with $400,000
  3. bPredictive ratio for a subgroup was computed as the ratio of the mean of observed to the mean of predicted spending over the subgroup. Approximate confidence intervals for predictive ratios were computed with the delta method [40]
  4. cNet compensation for a subgroup was computed as the mean difference between predicted and observed spending in the subgroup. Confidence intervals were estimated using a paired t-test