PSM table generation

cumulus_library/.sqlfluff

@@ -4,7 +4,7 @@ dialect = athena
 sql_file_exts = .sql,.sql.jinja
 # this rule overfires on athena nested arrays
 exclude_rules=references.from,structure.column_order,aliasing.unused
-max_line_length = 88
+max_line_length = 90
 
 [sqlfluff:indentation]
 template_blocks_indent = false
@@ -18,20 +18,30 @@ capitalisation_policy = upper
 [sqlfluff:templater:jinja:context]
 code_systems = ["http://snomed.info/sct", "http://hl7.org/fhir/sid/icd-10-cm"]
 col_type_list = ["a string","b string"]
+columns = ['a','b']
 cc_columns = [{"name": "baz", "is_array": True}, {"name": "foobar", "is_array": False}]
 cc_column = 'code'
 code_system_tables = [{table_name":"hasarray","column_name":"acol","is_bare_coding":False,"is_array":True, "has_data": True},{"table_name":"noarray","column_name":"col","is_bare_coding":False,"is_array":False, "has_data": True}{"table_name":"bare","column_name":"bcol","is_bare_coding":True,"is_array":False, "has_data": True},{"table_name":"empty","column_name":"empty","is_bare_coding":False,"is_array":False, "has_data": False}]
 column_name = 'bar'
 conditions = ["1 > 0", "1 < 2"]
+count_ref = count_ref
+count_table = count_table
 dataset = [["foo","foo"],["bar","bar"]]
+dependent_variable = is_flu
 ext_systems = ["omb", "text"]
 field = 'column_name'
+filter_table = filter_table
 fhir_extension = fhir_extension
 fhir_resource = patient
 id = 'id'
+join_cols_by_table = { "join_table": { "join_id": "enc_ref","included_cols": [["a"], ["b", "c"]]}}
+join_id = subject_ref
 medication_datasources = {"by_contained_ref" : True, "by_external_ref" : True}
+neg_source_table = neg_source_table
 output_table_name = 'created_table'
 prefix = Test
+primary_ref = encounter_ref
+pos_source_table = pos_source_table
 schema_name = test_schema
 source_table = source_table
 source_id = source_id

cumulus_library/base_table_builder.py

@@ -2,7 +2,6 @@
 import re
 
 from abc import ABC, abstractmethod
-from typing import final
 
 from cumulus_library.databases import DatabaseCursor
 from cumulus_library.helper import get_progress_bar, query_console_output
@@ -33,7 +32,12 @@ def prepare_queries(self, cursor: object, schema: str):
         """
         raise NotImplementedError
 
-    @final
+    # 🚨🚨🚨 WARNING: in 99% of cases, subclasses should *not* re-implement 🚨🚨🚨
+    # 🚨🚨🚨 execute_queries.                                               🚨🚨🚨
+
+    # If you know what you are doing, you can attempt to override it, but it is
+    # strongly recommended you invoke this as is via a super() call, and then
+    # run code before or after that as makes sense for your use case.
     def execute_queries(
         self,
         cursor: DatabaseCursor,

cumulus_library/databases.py

@@ -1,4 +1,12 @@
-"""Abstraction layers for supported database backends (e.g. AWS & DuckDB)"""
+"""Abstraction layers for supported database backends (e.g. AWS & DuckDB)
+
+By convention, to maintain this as a relatively light wrapper layer, if you have
+to chose between a convenience function in a specific library (as an example, the
+[pyathena to_sql function](https://github.com/laughingman7743/PyAthena/#to-sql))
+or using raw sql directly in some form, you should do the latter. This not a law;
+if there's a compelling reason to do so, just make sure you add an appropriate
+wrapper method in one of DatabaseCursor or DatabaseBackend.
+"""
 
 import abc
 import datetime
@@ -47,6 +55,14 @@ def __init__(self, schema_name: str):
     def cursor(self) -> DatabaseCursor:
         """Returns a connection to the backing database"""
 
+    @abc.abstractmethod
+    def pandas_cursor(self) -> DatabaseCursor:
+        """Returns a connection to the backing database optimized for dataframes
+
+        If your database does not provide an optimized cursor, this should function the
+        same as a vanilla cursor.
+        """
+
     @abc.abstractmethod
     def execute_as_pandas(self, sql: str) -> pandas.DataFrame:
         """Returns a pandas.DataFrame version of the results from the provided SQL"""
@@ -85,6 +101,9 @@ def __init__(self, region: str, workgroup: str, profile: str, schema_name: str):
     def cursor(self) -> AthenaCursor:
         return self.connection.cursor()
 
+    def pandas_cursor(self) -> AthenaPandasCursor:
+        return self.pandas_cursor
+
     def execute_as_pandas(self, sql: str) -> pandas.DataFrame:
         return self.pandas_cursor.execute(sql).as_pandas()
 
@@ -95,6 +114,8 @@ class DuckDatabaseBackend(DatabaseBackend):
     def __init__(self, db_file: str):
         super().__init__("main")
         self.connection = duckdb.connect(db_file)
+        # Aliasing Athena's as_pandas to duckDB's df cast
+        setattr(duckdb.DuckDBPyConnection, "as_pandas", duckdb.DuckDBPyConnection.df)
 
         # Paper over some syntax differences between Athena and DuckDB
         self.connection.create_function(
@@ -150,6 +171,10 @@ def cursor(self) -> duckdb.DuckDBPyConnection:
         # because then we'd have to re-register our json tables.
         return self.connection
 
+    def pandas_cursor(self) -> duckdb.DuckDBPyConnection:
+        # Since this is not provided, return the vanilla cursor
+        return self.connection
+
     def execute_as_pandas(self, sql: str) -> pandas.DataFrame:
         # We call convert_dtypes here in case there are integer columns.
         # Pandas will normally cast nullable-int as a float type unless

cumulus_library/statistics/psm.py

@@ -0,0 +1,247 @@
+# Module for generating Propensity Score matching cohorts
+
+import numpy as np
+import pandas
+import toml
+
+from psmpy import PsmPy
+
+
+import json
+from pathlib import PosixPath
+from dataclasses import dataclass
+
+from cumulus_library.cli import StudyBuilder
+from cumulus_library.databases import DatabaseCursor
+from cumulus_library.base_table_builder import BaseTableBuilder
+from cumulus_library.template_sql.templates import (
+    get_ctas_query_from_df,
+    get_drop_view_table,
+)
+from cumulus_library.template_sql.statistics.psm_templates import (
+    get_distinct_ids,
+    get_create_covariate_table,
+)
+
+
+@dataclass
+class PsmConfig:
+    """Provides expected values for PSM execution
+
+    These values should be read in from a toml configuration file.
+    See tests/test_data/psm/psm_config.toml for an example with details about
+    the expected values for these fields.
+    """
+
+    classification_json: str
+    pos_source_table: str
+    neg_source_table: str
+    target_table: str
+    primary_ref: str
+    count_ref: str
+    count_table: str
+    dependent_variable: str
+    pos_sample_size: int
+    neg_sample_size: int
+    join_cols_by_table: dict[str, dict]
+    seed: int
+
+
+class PsmBuilder(BaseTableBuilder):
+    """TableBuilder for creating PSM tables"""
+
+    display_text = "Building PSM tables..."
+
+    def __init__(self, toml_config_path: str):
+        """Loads PSM job details from a PSM configuration file"""
+        with open(toml_config_path, encoding="UTF-8") as file:
+            toml_config = toml.load(file)
+        self.config = PsmConfig(
+            classification_json=f"{PosixPath(toml_config_path).parent}/{toml_config['classification_json']}",
+            pos_source_table=toml_config["pos_source_table"],
+            neg_source_table=toml_config["neg_source_table"],
+            target_table=toml_config["target_table"],
+            primary_ref=toml_config["primary_ref"],
+            dependent_variable=toml_config["dependent_variable"],
+            pos_sample_size=toml_config["pos_sample_size"],
+            neg_sample_size=toml_config["neg_sample_size"],
+            join_cols_by_table=toml_config.get("join_cols_by_table", {}),
+            count_ref=toml_config.get("count_ref", None),
+            count_table=toml_config.get("count_table", None),
+            seed=toml_config.get("seed", 123),
+        )
+        super().__init__()
+
+    def _get_symptoms_dict(self, path: str) -> dict:
+        """convenience function for loading symptoms dictionaries from a json file"""
+        with open(path) as f:
+            symptoms = json.load(f)
+        return symptoms
+
+    def _get_sampled_ids(
+        self,
+        cursor: DatabaseCursor,
+        schema: str,
+        query: str,
+        sample_size: int,
+        dependent_variable: str,
+        is_positive: bool,
+    ):
+        """Creates a table containing randomly sampled patients for PSM analysis
+
+        To use this, it is assumed you have already identified a cohort of positively
+        IDed patients as a manual process.
+        :param cursor: A valid DatabaseCusror:
+        :param schema: the schema/database name where the data exists
+        :param query: a query generated from the psm_dsitinct_ids template
+        :param sample_size: the number of records to include in the random sample.
+            This should generally be >= 20.
+        :param dependent_variable: the name to use for your filtering column
+        :param is_positive: defines the value to be used for your filtering column
+        """
+        df = cursor.execute(query).as_pandas()
+        df = (
+            df.sort_values(by=[self.config.primary_ref])
+            .reset_index()
+            .drop("index", axis=1)
+        )
+        df = (
+            # TODO: remove replace behavior after increasing data sample size
+            df.sample(n=sample_size, random_state=self.config.seed, replace=True)
+            .sort_values(by=[self.config.primary_ref])
+            .reset_index()
+            .drop("index", axis=1)
+        )
+
+        df[dependent_variable] = is_positive
+        return df
+
+    def _create_covariate_table(self, cursor: DatabaseCursor, schema: str):
+        """Creates a covariate table from the loaded toml config"""
+        # checks for primary & link ref being the same
+        source_refs = list(
+            {self.config.primary_ref, self.config.count_ref} - set([None])
+        )
+        pos_query = get_distinct_ids(source_refs, self.config.pos_source_table)
+        pos = self._get_sampled_ids(
+            cursor,
+            schema,
+            pos_query,
+            self.config.pos_sample_size,
+            self.config.dependent_variable,
+            1,
+        )
+        neg_query = get_distinct_ids(
+            source_refs,
+            self.config.neg_source_table,
+            join_id=self.config.primary_ref,
+            filter_table=self.config.pos_source_table,
+        )
+        neg = self._get_sampled_ids(
+            cursor,
+            schema,
+            neg_query,
+            self.config.neg_sample_size,
+            self.config.dependent_variable,
+            0,
+        )
+        cohort = pandas.concat([pos, neg])
+
+        # Replace table (if it exists)
+        # TODO - replace with timestamp prepended table
+        drop = get_drop_view_table(
+            f"{self.config.pos_source_table}_sampled_ids", "TABLE"
+        )
+        cursor.execute(drop)
+        ctas_query = get_ctas_query_from_df(
+            schema,
+            f"{self.config.pos_source_table}_sampled_ids",
+            cohort,
+        )
+        self.queries.append(ctas_query)
+        # TODO - replace with timestamp prepended table
+        drop = get_drop_view_table(self.config.target_table, "TABLE")
+        cursor.execute(drop)
+        dataset_query = get_create_covariate_table(
+            target_table=self.config.target_table,
+            pos_source_table=self.config.pos_source_table,
+            neg_source_table=self.config.neg_source_table,
+            primary_ref=self.config.primary_ref,
+            dependent_variable=self.config.dependent_variable,
+            join_cols_by_table=self.config.join_cols_by_table,
+            count_ref=self.config.count_ref,
+            count_table=self.config.count_table,
+        )
+        self.queries.append(dataset_query)
+
+    def generate_psm_analysis(self, cursor: object, schema: str):
+        """Runs PSM statistics on generated tables"""
+        df = cursor.execute(f"select * from {self.config.target_table}").as_pandas()
+        symptoms_dict = self._get_symptoms_dict(self.config.classification_json)
+        for dependent_variable, codes in symptoms_dict.items():
+            df[dependent_variable] = df["code"].apply(lambda x: 1 if x in codes else 0)
+        df = df.drop(columns="code")
+        # instance_count present but unused for PSM if table contains a count_ref input
+        # (it's intended for manual review)
+        df = df.drop(columns="instance_count", errors="ignore")
+
+        columns = []
+        if self.config.join_cols_by_table is not None:
+            for table_key in self.config.join_cols_by_table:
+                for column in self.config.join_cols_by_table[table_key][
+                    "included_cols"
+                ]:
+                    if len(column) == 2:
+                        columns.append(column[1])
+                    else:
+                        columns.append(column[0])
+
+        for column in columns:
+            encoded_df = pandas.get_dummies(df[column])
+            df = pandas.concat([df, encoded_df], axis=1)
+            df = df.drop(column, axis=1)
+        df = df.reset_index()
+
+        try:
+            psm = PsmPy(
+                df,
+                treatment=self.config.dependent_variable,
+                indx=self.config.primary_ref,
+                exclude=[],
+            )
+            # This function populates the psm.predicted_data element, which is required
+            # for things like the knn_matched() function call
+            psm.logistic_ps(balance=True)
+            print(psm.predicted_data)
+            # This function populates the psm.df_matched element
+            # TODO: flip replacement to false after increasing sample data size
+            psm.knn_matched(
+                matcher="propensity_logit",
+                replacement=True,
+                caliper=None,
+                drop_unmatched=True,
+            )
+            print(psm.df_matched)
+        except ZeroDivisionError:
+            print(
+                "Encountered a divide by zero error during statistical graph generation. Try increasing your sample size."
+            )
+        except ValueError:
+            print(
+                "Encountered a value error during KNN matching. Try increasing your sample size."
+            )
+
+    def prepare_queries(self, cursor: object, schema: str):
+        self._create_covariate_table(cursor, schema)
+
+    def execute_queries(
+        self,
+        cursor: object,
+        schema: str,
+        verbose: bool,
+        drop_table: bool = False,
+    ):
+        super().execute_queries(cursor, schema, verbose, drop_table)
+        self.comment_queries()
+        self.write_queries()
+        self.generate_psm_analysis(cursor, schema)