Canonical text (the exact bytes fingerprinted)

TRUVACE RECORD VERSION
record: TRV-2026-1482
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-10-15T18:14:59.198445Z
status: published
lens: trace
sector: health
headline: Utilizing categorical boosting and SHAP to understand key predictors of frequency of use at discharge for completed substance abuse treatments in TEDS-D
dek: Background: Substance use disorder is a pressing US public health crisis, with 48.7 million people reporting past-year substance use and over 100 000 overdose deaths in 2022. Building on a growing body of machine learning research using the SAMHSA treatment episode Data set-discharges (TEDS-D), this study introduces frequency of use at discharge as an outcome variable and the inclusion of U.S. state as a predictor. Research Design and Methods: First, reason for discharge was evaluated as a dependent variable and…
gain_title: A boosted-tree model with SHAP explanations predicted substance use frequency at discharge with strong accuracy and showed that treatment and geographic factors, not patient attributes, mattered most, giving a basis for tailoring treatment.
problem_title: State-level models agreed only moderately with the national model, and federal data substantially underreported some treatment modalities in Hawai'i, which complicates national benchmarking and may obscure true drivers of recovery.
trace_subject: Machine learning prediction of substance use at discharge from completed treatment in TEDS-D
gain_reading: A boosted-tree model with SHAP explanations predicted substance use frequency at discharge with strong accuracy and showed that treatment and geographic factors, not patient attributes, mattered most, giving a basis for tailoring treatment.
gain_evidence: Treatment and geographic factors outweighed patient attributes
problem_reading: State-level models agreed only moderately with the national model, and federal data substantially underreported some treatment modalities in Hawai'i, which complicates national benchmarking and may obscure true drivers of recovery.
problem_evidence: revealed substantial underreporting of long-term residential and therapeutic community treatment modalities in federal data
quick_read: Researchers trained a categorical boosting model on about 2.4 million completed treatment episodes from the 2015-2019 national discharge data set to predict how often clients used substances at discharge. They used SHAP to rank which of 62 variables mattered most and compared state-specific models with the national one.

The work matters because completing treatment did not guarantee good outcomes, and the findings suggest that where and how treatment is delivered shapes results. Reporting inconsistencies across states could distort national comparisons, and the analysis is based on administrative data from 2015-2019, so its use for current decisions remains to be shown.
limitation: State-level models showed only moderate agreement with the national model, and underreporting in federal data means national benchmarks may not reflect the true drivers of recovery.
tag: Dual reading
key_points: Treatment completion gave higher no-use rates at discharge than dropout (45% vs 28%). | Among completed episodes, 25% reported active substance use, 9% homelessness and 50% unemployment at discharge. | The four most important features were use frequency at admission, state of treatment, length of stay and service type. | Long-term residential rehabilitation had the lowest SHAP values and lowest predicted probability of use at discharge. | The Hawai'i case study found elevated predicted probability of use at discharge for younger clients.
rundown: The study used 2,411,351 completed treatment episodes from 2015-2019 and 62 predictor variables. It first compared reason for discharge with frequency of use at discharge as the outcome, then built the main prediction model around frequency of use.

SHAP supplied both global feature importance and explanations of individual predictions. State-level models were compared with the national model through rank-biased overlap, which averaged 0.59 and ranged from 0.30 to 0.73.

A Hawai'i case study showed how state-level decision support could work. It highlighted younger clients at higher risk and found that long-term residential and therapeutic community treatment were substantially underreported in federal data.
sources:
- peer_reviewed | Machine Learning: Health | https://doi.org/10.1088/3049-477x/ae9bc7 | 2026-10-07
prev: 0000000000000000000000000000000000000000000000000000000000000000

sha256

49f5b61f532fc80829a4c606c1fa94e4d803c04f0b30601a3c0264e52b582c57

previous

0000000000000000000000000000000000000000000000000000000000000000