Risk scores are sparse linear models that map an integer linear combination of covariates to the probability of an outcome occurring. Unlike regression models, risk score models consist of integer coefficients for often dichotomous variables. This allows risk score predictions to be easily computed by adding or subtracting a few small numbers.
Risk scores developed heuristically by altering logistic regression
models have decreased performance, as there is a fundamental trade-off
between the model’s simplicity and its predictive accuracy. In contrast,
this package presents an optimization approach to learning risk scores,
where the constraints unique to risk score models are integrated into
the model-fitting process, rather than implemented afterward. This
vignette demonstrates how to use the riskscores
package to
build a risk score model to predict breast cancer diagnosis.
The riskscores
package uses a cyclical coordinate
descent algorithm to solve the following optimization problem.
\[\begin{equation} \begin{aligned} \min_{\alpha,\beta} \quad & \frac{1}{n} \sum_{i=1}^{n} (\gamma y_i x_i^T \beta - log(1 + exp(\gamma x_i^T \beta))) + \lambda_0 \sum_{j=1}^{p} 1(\beta_{j} \neq 0)\\ \textrm{s.t.} \quad & l \le \beta_j \le u \; \; \; \forall j = 1,2,...,p\\ &\beta_j \in \mathbb{Z} \; \; \; \forall j = 1,2,...,p \\ &\beta_0, \gamma \in \mathbb{R} \\ \end{aligned} \end{equation}\]
These constraints ensure that the model will be sparse and include only integer coefficients.
First we’ll load in an example dataset. In this example, we want to
develop a risk score model that predicts whether a breast tissue sample
is benign using features recorded during a biopsy. The
breastcancer
dataset was originally accessed from the UCI
Repository and can be loaded into your environment from the
riskscores
package as so:
This dataset contains 683 observations and 9 features. Our goal is to
develop a risk score model that predicts whether a breast tissue sample
is benign using 9 (or fewer) features recorded during a biopsy:
Before building a risk score model, data often need to be preprocessed. Specifically, the dataset needs to have a binary outcome with all other variables containing either binary or integer values.
The breastcancer
dataset is mostly ready to go. We’ll
still need to split out our data into a matrix with all covariates
(X
) and a vector with the outcome data (y
). In
this case, the first column in our dataset contains the outcome
variable.
The penalty coefficient \(\lambda_0\) controls the sparsity of the model – a larger value of \(\lambda_0\) will result in fewer non-zero coefficients. We can use cross validation to find the optimal \(\lambda_0\) value that creates a sufficiently sparse model without sacrificing performance.
Ideally, each cross-validation fold should contain an approximately
equal proportion of cases. The riskscores
package contains
the function stratify_folds()
that creates fold IDs with an
equal proportion of cases in each fold. These fold IDs can be entered
into the cv_risk_mod()
function under the
foldids
parameter. Otherwise, cv_risk_mod()
will set random fold IDs.
The cv_risk_mod()
function runs cross validation for a
grid of possible \(\lambda_0\) values.
If the user does not specify the vector of \(\lambda_0\) values to test, the program
constructs this \(\lambda_0\) sequence.
The maximum \(\lambda_0\) in this
sequence is the smallest value such that all coefficients in the
logistic regression model are zero. The minimum \(\lambda_0\) in the sequence is calculated
using the user-defined lambda_ratio
argument. The \(\lambda_0\) grid is created by generating
nlambda
values linear on the log scale from the minimum
\(\lambda_0\) to the maximum \(\lambda_0\). We’ve set nlambda
to 25, so the program will construct an appropriate sequence of 25 \(\lambda_0\) values to test using cross
validation.
Running plot()
on a cv_risk_mod
object
creates a plot of mean deviance for each \(\lambda_0\) value in the grid. The number
of nonzero coefficients that are produced by each \(\lambda_0\) value when fit on the full data
are listed at the top of the plot. The \(\lambda_0\) value with the lowest mean
deviance (“lambda_min”) is indicated in red, and its standard deviation
is marked with a red dashed line. Its precise value can be accessed by
calling cv_results$lambda_min
. If we want a sparser model,
we could increase \(\lambda_0\) to
“lambda_1se”, the largest value whose mean deviance is within one
standard error of “lambda_min”. This value can be accessed by calling
cv_results$lambda_1se
. In our example, “lambda_min” creates
a model with 8 non-zero coefficients and “lambda_1se” creates a model
with 3 non-zero coefficients.
To view a dataframe with the full cross-validation results (including
both deviance and accuracy metrics), run
cv_results$results
.
tail(cv_results$results)
#> # A tibble: 6 × 6
#> lambda0 mean_dev sd_dev mean_acc sd_acc nonzero
#> <dbl> <dbl> <dbl> <dbl> <dbl> <int>
#> 1 0.0576 33.1 10.8 0.956 0.0103 3
#> 2 0.0845 50.2 20.2 0.934 0.0313 3
#> 3 0.124 47.8 18.4 0.930 0.0239 2
#> 4 0.182 53.7 19.7 0.930 0.0285 2
#> 5 0.267 49.2 23.1 0.931 0.0210 1
#> 6 0.392 81.4 37.9 0.884 0.0650 1
We’ll fit a model on the full data using the function
risk_mod()
. We’ll use the “lambda_1se” value determined by
cross-validation as our \(\lambda_0\)
parameter.
The integer risk score model can be viewed by calling
mod$model_card
. An individual’s risk score can be
calculated by multiplying each covariate response by its respective
number of points and then adding all points together. In our example
below, a patient with a ClumpThickness value of 1, a
UniformityOfCellShape value of 10, and a BareNuclei value of 5 would
receive a score of \(10(1) + 8(10) + 7(5) =
125\).
Points | |
---|---|
ClumpThickness | 10 |
UniformityOfCellShape | 8 |
BareNuclei | 7 |
Each score can then be mapped to a risk probability. The
mod$score_map
dataframe maps an integer range of scores to
their associated risk. For this example dataset,
mod$score_map
includes a range of integer scores from 25 to
200, which are the minimum and maximum scores predicted from the
training data. The table below shows a sample of these scores mapped to
their associated risk. We can see that a patient who received a score of
125 would have a 77.65% risk of their tissue sample being malignant.
Score | Risk |
---|---|
25 | 0.0014 |
50 | 0.0096 |
75 | 0.0644 |
100 | 0.3284 |
125 | 0.7765 |
150 | 0.9611 |
175 | 0.9943 |
200 | 0.9992 |
The function get_risk()
can be used to calculate the
risk from a given score (or a vector of scores). Likewise, the function
get_score()
calculates the score associated with a given
risk (or vector of risk probabilities).
We can evaluate the model’s performance under different
classification thresholds using the get_metrics()
function.
get_metrics(mod, threshold = seq(0.1, 0.9, 0.1))
#> threshold_risk threshold_score accuracy sensitivity specificity
#> 1 0.1 81.1 0.9546120 0.9958159 0.9324324
#> 2 0.2 91.4 0.9648609 0.9748954 0.9594595
#> 3 0.3 98.3 0.9677892 0.9581590 0.9729730
#> 4 0.4 103.9 0.9677892 0.9539749 0.9752252
#> 5 0.5 109.1 0.9648609 0.9372385 0.9797297
#> 6 0.6 114.3 0.9619327 0.9288703 0.9797297
#> 7 0.7 119.9 0.9590044 0.9121339 0.9842342
#> 8 0.8 126.8 0.9458272 0.8744770 0.9842342
#> 9 0.9 137.1 0.9326501 0.8326360 0.9864865
Running summary()
on our model will return the
intercept, the scores of each nonzero coefficient, the \(\gamma\) multiplier value, the \(\lambda_0\) regularizer value, the
deviance, and the AIC.
A vector containing the risk score model intercept and integer
coefficients can be accessed by calling coef()
on the
risk_mod
object. This vector is also saved as
$beta
within the risk_mod
object.
coef(mod) # equivalently: mod$beta
#> Intercept ClumpThickness UniformityOfCellSize
#> -109.1192 10.0000 0.0000
#> UniformityOfCellShape MarginalAdhesion SingleEpithelialCellSize
#> 8.0000 0.0000 0.0000
#> BareNuclei BlandChromatin NormalNucleoli
#> 7.0000 0.0000 0.0000
#> Mitoses
#> 0.0000
We can map our integer score model to an equivalent logistic
regression model by multiplying the integer and coefficients by \(\gamma\) (saved as $gamma
in
the risk_mod
object).
coef(mod) * mod$gamma
#> Intercept ClumpThickness UniformityOfCellSize
#> -8.5584971 0.7843255 0.0000000
#> UniformityOfCellShape MarginalAdhesion SingleEpithelialCellSize
#> 0.6274604 0.0000000 0.0000000
#> BareNuclei BlandChromatin NormalNucleoli
#> 0.5490279 0.0000000 0.0000000
#> Mitoses
#> 0.0000000
The risk_mod
object stores a glm
object of
this non-integer logistic regression model as $glm_mod
.
coef(mod$glm_mod)
#> Intercept ClumpThickness UniformityOfCellSize
#> -8.5584971 0.7843255 0.0000000
#> UniformityOfCellShape MarginalAdhesion SingleEpithelialCellSize
#> 0.6274604 0.0000000 0.0000000
#> BareNuclei BlandChromatin NormalNucleoli
#> 0.5490279 0.0000000 0.0000000
#> Mitoses
#> 0.0000000
Running predict()
on a risk_mod
object
allows for three types of prediction, as the type
parameter
can be set to either 'link'
, 'response'
, or
'score'
. These first two options are the same as when
predict()
is run on a logistic glm
object. The
added 'score'
option returns each subject’s score, as
calculated from the integer coefficients in the risk score model.
The table below compares the three possible prediction types for five example subjects. The first three columns contain data for clump thickness, uniformity of cell shape, and bare nuclei, respectively.
CT | UCS | BN | ‘score’ | ‘link’ | ‘response’ |
---|---|---|---|---|---|
5 | 1 | 1 | 65 | -3.46 | 0.030 |
5 | 4 | 10 | 152 | 3.36 | 0.967 |
3 | 1 | 2 | 52 | -4.48 | 0.011 |
6 | 8 | 4 | 152 | 3.36 | 0.967 |
4 | 1 | 1 | 55 | -4.24 | 0.014 |
The ‘score’ is a linear combination of the covariates and their integer coefficients:
The ‘link’ is a linear combination of the covariates using the full logistic regression equation:
The ‘response’ converts these link values to probabilities: