Add canonical PerformanceTableSpec producers

src/rtichoke/_performance_table_spec.py

@@ -0,0 +1,222 @@
+"""Internal canonical PerformanceTableSpec builders.
+
+These helpers translate already-calculated production performance data plus
+semantic evaluation metadata. They do not calculate or render statistics.
+"""
+
+from __future__ import annotations
+
+import math
+from collections.abc import Mapping
+from typing import Any, TypedDict, cast
+
+import polars as pl
+
+from rtichoke.processing.evaluation_semantics import _EvaluationMetadata
+
+_METRICS: tuple[tuple[str, str], ...] = (
+    ("true_positives", "True Positives"),
+    ("true_negatives", "True Negatives"),
+    ("false_positives", "False Positives"),
+    ("false_negatives", "False Negatives"),
+    ("sensitivity", "Sensitivity"),
+    ("specificity", "Specificity"),
+    ("false_positive_rate", "False Positive Rate"),
+    ("ppv", "PPV"),
+    ("npv", "NPV"),
+    ("lift", "Lift"),
+    ("predicted_positives", "Predicted Positives"),
+    ("ppcr", "PPCR"),
+    ("net_benefit", "Net Benefit"),
+    ("net_benefit_interventions_avoided", "Interventions Avoided"),
+)
+
+
+class _EvaluationSpec(TypedDict, total=False):
+    id: str
+    model: str
+    population: str
+
+
+class _MetricDefinition(TypedDict):
+    id: str
+    label: str
+
+
+class _OperatingPoint(TypedDict):
+    type: str
+    value: float
+
+
+class _MetricValue(TypedDict):
+    metricId: str
+    estimate: float | int | None
+
+
+class _EvaluationContext(TypedDict):
+    censoringHeuristic: str
+    competingEventHeuristic: str
+
+
+class _PerformanceTableRow(TypedDict, total=False):
+    evaluationId: str
+    operatingPoint: _OperatingPoint
+    values: list[_MetricValue]
+    horizon: float
+    context: _EvaluationContext
+
+
+class _PerformanceTableSpec(TypedDict):
+    schemaVersion: str
+    type: str
+    evaluations: list[_EvaluationSpec]
+    metrics: list[_MetricDefinition]
+    rows: list[_PerformanceTableRow]
+
+
+def _performance_table_spec_from_performance_data(
+    performance_data: pl.DataFrame,
+    evaluation_metadata: Mapping[str, _EvaluationMetadata],
+) -> _PerformanceTableSpec:
+    """Build the canonical static PerformanceTableSpec."""
+    return _build_performance_table_spec(
+        performance_data, evaluation_metadata, time_dependent=False
+    )
+
+
+def _performance_table_times_spec_from_performance_data(
+    performance_data: pl.DataFrame,
+    evaluation_metadata: Mapping[str, _EvaluationMetadata],
+) -> _PerformanceTableSpec:
+    """Build the canonical time-dependent PerformanceTableSpec."""
+    return _build_performance_table_spec(
+        performance_data, evaluation_metadata, time_dependent=True
+    )
+
+
+def _build_performance_table_spec(
+    performance_data: pl.DataFrame,
+    evaluation_metadata: Mapping[str, _EvaluationMetadata],
+    *,
+    time_dependent: bool,
+) -> _PerformanceTableSpec:
+    required = {"reference_group", "stratified_by", "chosen_cutoff"}
+    if time_dependent:
+        required |= {
+            "fixed_time_horizon",
+            "censoring_heuristic",
+            "competing_heuristic",
+        }
+    missing = required.difference(performance_data.columns)
+    if missing:
+        raise ValueError(
+            "Performance-table data is missing columns: " + ", ".join(sorted(missing))
+        )
+
+    metric_definitions: list[_MetricDefinition] = [
+        cast(_MetricDefinition, {"id": metric_id, "label": label})
+        for metric_id, label in _METRICS
+        if metric_id in performance_data.columns
+    ]
+    metric_ids = [definition["id"] for definition in metric_definitions]
+
+    rows = performance_data.select(
+        [
+            column
+            for column in (
+                "reference_group",
+                "fixed_time_horizon",
+                "censoring_heuristic",
+                "competing_heuristic",
+                "stratified_by",
+                "chosen_cutoff",
+                *metric_ids,
+            )
+            if column in performance_data.columns
+        ]
+    ).to_dicts()
+    row_groups = {str(row["reference_group"]) for row in rows}
+    missing_metadata = row_groups.difference(evaluation_metadata)
+    if missing_metadata:
+        raise ValueError(
+            "Performance-table rows are missing evaluation metadata: "
+            + ", ".join(sorted(missing_metadata))
+        )
+
+    ordered_groups = [group for group in evaluation_metadata if group in row_groups]
+    evaluation_ids = {
+        group: f"evaluation-{index}"
+        for index, group in enumerate(ordered_groups, start=1)
+    }
+    evaluations: list[_EvaluationSpec] = []
+    for group in ordered_groups:
+        metadata = evaluation_metadata[group]
+        evaluation: _EvaluationSpec = {
+            "id": evaluation_ids[group],
+            "population": metadata.population,
+        }
+        if metadata.model is not None:
+            evaluation["model"] = metadata.model
+        evaluations.append(evaluation)
+
+    canonical_rows: list[_PerformanceTableRow] = []
+    for row in rows:
+        group = str(row["reference_group"])
+        stratified_by = str(row["stratified_by"])
+        operating_point: _OperatingPoint
+        if stratified_by == "probability_threshold":
+            operating_point = {
+                "type": "probability_threshold",
+                "value": _number(row["chosen_cutoff"]),
+            }
+        elif stratified_by == "ppcr":
+            operating_point = {"type": "ppcr", "value": _number(row["ppcr"])}
+        else:
+            raise ValueError(
+                "Canonical PerformanceTableSpec supports probability_threshold "
+                f"or ppcr operating points, not {stratified_by!r}"
+            )
+
+        canonical_row: _PerformanceTableRow = {
+            "evaluationId": evaluation_ids[group],
+            "operatingPoint": operating_point,
+            "values": cast(
+                list[_MetricValue],
+                [
+                    {
+                        "metricId": metric_id,
+                        "estimate": _nullable_number(row[metric_id]),
+                    }
+                    for metric_id in metric_ids
+                ],
+            ),
+        }
+        if time_dependent:
+            canonical_row["horizon"] = _number(row["fixed_time_horizon"])
+            canonical_row["context"] = {
+                "censoringHeuristic": str(row["censoring_heuristic"]),
+                "competingEventHeuristic": str(row["competing_heuristic"]),
+            }
+        canonical_rows.append(canonical_row)
+
+    return {
+        "schemaVersion": "2.0",
+        "type": "performance_table",
+        "evaluations": evaluations,
+        "metrics": metric_definitions,
+        "rows": canonical_rows,
+    }
+
+
+def _nullable_number(value: Any) -> float | int | None:
+    if value is None:
+        return None
+    if isinstance(value, float) and math.isnan(value):
+        return None
+    return value
+
+
+def _number(value: Any) -> float:
+    if value is None:
+        raise ValueError("Operating point and horizon values must not be null")
+    return float(value)

src/rtichoke/processing/adjustments.py

@@ -1,8 +1,9 @@
+from collections.abc import Sequence
+
 import pandas as pd
 import polars as pl
-from polarstate import predict_aj_estimates
-from polarstate import prepare_event_table
-from collections.abc import Sequence
+from polarstate import predict_aj_estimates, prepare_event_table
+
 from rtichoke.processing.transforms import assign_and_explode_polars
 
 
@@ -529,7 +530,7 @@ def _aj_estimates_by_cutoff_per_horizon(
             df.filter(pl.col("fixed_time_horizon") == h)
             .group_by("strata")
             .map_groups(
-                lambda group: extract_aj_estimate_by_cutoffs(
+                lambda group, h=h: extract_aj_estimate_by_cutoffs(
                     group, [h], breaks, stratified_by, full_event_table=False
                 )
             )
@@ -547,7 +548,7 @@ def _aj_estimates_per_horizon(
             df.filter(pl.col("fixed_time_horizon") == h)
             .group_by("strata")
             .map_groups(
-                lambda group: extract_aj_estimate_for_strata(
+                lambda group, h=h: extract_aj_estimate_for_strata(
                     group, [h], full_event_table
                 )
             )
@@ -650,6 +651,8 @@ def _aj_adjusted_events(
             adjusted = extract_aj_estimate_by_cutoffs(
                 non_competing, horizons, breaks, stratified_by, full_event_table
             )
+        else:
+            raise ValueError(f"Unsupported risk-set scope: {risk_set_scope!r}")
 
         adjusted = adjusted.with_columns(
             [
@@ -704,6 +707,8 @@ def _aj_adjusted_events(
             adjusted = _aj_estimates_by_cutoff_per_horizon(
                 base_df, horizons, breaks, stratified_by
             )
+        else:
+            raise ValueError(f"Unsupported risk-set scope: {risk_set_scope!r}")
 
         adjusted = adjusted.with_columns(
             pl.lit(risk_set_scope)
@@ -730,6 +735,8 @@ def _aj_adjusted_events(
             adjusted = extract_aj_estimate_by_cutoffs(
                 base_df, horizons, breaks, stratified_by, full_event_table
             )
+        else:
+            raise ValueError(f"Unsupported risk-set scope: {risk_set_scope!r}")
 
         adjusted = adjusted.with_columns(
             [

tests/test_cutoff_grid_endpoint.py

@@ -1,3 +1,5 @@
+from typing import cast
+
 import numpy as np
 import pytest
 
@@ -17,7 +19,7 @@ def test_probability_threshold_breaks_match_r_seq_semantics(by, expected):
     breaks = create_breaks_values(None, "probability_threshold", by)
 
     assert breaks.tolist() == expected
-    assert np.max(breaks) <= 1.0
+    assert float(np.max(breaks)) <= 1.0
 
 
 @pytest.mark.parametrize(
@@ -33,7 +35,7 @@ def test_binary_performance_data_matches_r_cutoff_endpoint(by, expected_max):
 
     assert cutoffs.min() == 0.0
     assert cutoffs.max() == expected_max
-    assert cutoffs.max() <= 1.0
+    assert cast(float, cutoffs.max()) <= 1.0
 
 
 @pytest.mark.parametrize(
@@ -56,4 +58,4 @@ def test_time_performance_data_matches_r_cutoff_endpoint(by, expected_max):
 
     assert cutoffs.min() == 0.0
     assert cutoffs.max() == expected_max
-    assert cutoffs.max() <= 1.0
+    assert cast(float, cutoffs.max()) <= 1.0

tests/test_performance_table.py

@@ -1,18 +1,20 @@
+from typing import cast
+
 import numpy as np
 import pytest
 from great_tables import GT
 from reactable import Reactable
 
 import rtichoke.performance_table as performance_table_module
-from rtichoke.performance_table_reactable import _bar_style, _net_benefit_style
-
 from rtichoke import (
     create_performance_table,
     create_performance_table_times,
     prepare_performance_data,
     prepare_performance_data_times,
     render_performance_table,
 )
+from rtichoke.performance_table import PerformanceTableRenderer
+from rtichoke.performance_table_reactable import _bar_style, _net_benefit_style
 
 
 def _example():
@@ -215,7 +217,9 @@ def test_invalid_renderer_is_rejected():
     probs, reals = _example()
     data = prepare_performance_data(probs, reals, by=0.1)
     with pytest.raises(ValueError, match="renderer"):
-        render_performance_table(data, renderer="unknown")
+        render_performance_table(
+            data, renderer=cast(PerformanceTableRenderer, "unknown")
+        )
 
 
 def test_reactable_metric_bar_matches_r_colors_and_geometry():