Skip to contents

Score every candidate symptom combination

Source: R/score_all_combinations.R
score_all_combinations.Rd

Scores every candidate combination of n_symptoms PCL-5 items against the full DSM-5-TR diagnosis and returns the complete ranked table – not just the best ones. This is the exhaustive companion to optimize_combinations / optimize_combinations_clusters (which keep only the top n_top): use it to study how performance decays across the whole candidate set, e.g. to show that many symptom sets are near-interchangeable (a plateau of near-optimal combinations followed by a drop).

Usage

score_all_combinations(
  data,
  n_symptoms = 6,
  n_required = 4,
  clusters = NULL,
  score_by = "balanced_accuracy",
  chunk_size = 1000,
  show_progress = TRUE
)

Arguments

data

A dataframe with the 20 PCL-5 item columns symptom_1 through symptom_20 (output of rename_ptsd_columns). Additional carry-through columns are ignored.

n_symptoms

Integer. Number of items per combination (default 6).

n_required

Integer. How many of the items must be present (score >= 2) for a positive diagnosis (default 4).

clusters

NULL (default) to score all subsets without a cluster constraint, or a named list of integer vectors (e.g. list(B = 1:5, C = 6:7, D = 8:14, E = 15:20)) for the cluster-constrained candidate set and diagnosis rule.

score_by

Character. Metric that defines the ranking: "balanced_accuracy" (default), "accuracy", or "sensitivity". All metrics are returned regardless; this only sets the sort order.

chunk_size

Integer. Number of combinations scored per chunk (default 1000). Affects speed and parallel granularity only, never the result.

show_progress

Logical. If TRUE (default), display a progress bar (sequential mode only).

Value

A data.frame with one row per candidate combination, sorted best-to-worst by score_by (ties broken by combination_id for determinism):

  • rank: 1 = best.

  • combination_id: sorted item numbers joined by underscores (e.g. "1_6_8_10_15_19").

  • tp, fn, fp, tn: the 2x2 counts against the full DSM-5-TR diagnosis.

  • sensitivity, specificity, ppv, npv, accuracy, balanced_accuracy: metrics on the 0-1 scale (NA where a denominator is zero).

The attributes n_symptoms, n_required, clusters, score_by, and n_combinations record the configuration.

Details

With clusters = NULL, all choose(20, n_symptoms) subsets are scored (38,760 for six symptoms). With a cluster structure, only the combinations containing at least one item per cluster are scored (13,685 six-symptom sets for the default PCL-5 clusters), and the diagnosis additionally requires the present symptoms to span all clusters – the same candidate set and rule as optimize_combinations_clusters. The hierarchical per-row cluster check makes this mode noticeably slower.

Combinations are processed in chunks. If the future.apply package is installed and a future plan is set (e.g. future::plan(future::multisession)), chunks are scored in parallel; otherwise they are scored sequentially with a progress bar. Results are identical either way.

The returned combination_id uses the same canonical format as write_combinations (sorted item numbers joined by underscores), so the full curve can be joined against exported top-k combinations.

See also

optimize_combinations, optimize_combinations_clusters, compare_optimizations.

Examples

# \donttest{
# A 4-symptom search on a 250-row subset keeps the example fast
ptsd_data <- rename_ptsd_columns(simulated_ptsd[1:250, ],
                                 id_col = c("patient_id", "age", "sex"))
curve <- score_all_combinations(ptsd_data, n_symptoms = 4, n_required = 3,
                                show_progress = FALSE)
nrow(curve)     # choose(20, 4) = 4845 combinations, all ranked
#> [1] 4845
head(curve)
#>   rank combination_id  tp fn fp tn sensitivity specificity ppv       npv
#> 1    1      6_7_12_17 227  5  0 18   0.9784483           1   1 0.7826087
#> 2    2       4_6_7_12 226  6  0 18   0.9741379           1   1 0.7500000
#> 3    3       4_6_7_19 225  7  0 18   0.9698276           1   1 0.7200000
#> 4    4      6_7_12_13 225  7  0 18   0.9698276           1   1 0.7200000
#> 5    5      6_7_12_15 225  7  0 18   0.9698276           1   1 0.7200000
#> 6    6      6_7_12_18 225  7  0 18   0.9698276           1   1 0.7200000
#>   accuracy balanced_accuracy
#> 1    0.980         0.9892241
#> 2    0.976         0.9870690
#> 3    0.972         0.9849138
#> 4    0.972         0.9849138
#> 5    0.972         0.9849138
#> 6    0.972         0.9849138

# The full balanced-accuracy curve, e.g. for a rank plot:
# plot(curve$rank, curve$balanced_accuracy, type = "l", log = "x")
# }