Visualizing the Right Counseling Support: Evidence-Linked Recommendation Cards for Explainable Mental Health Intake Interfaces

Abstract

This study presents and evaluates a reproducible interface pipeline for turning mental-health counseling case material into structured, evidence-linked decision-support cards. The aim is not to replace therapists or to claim clinical effectiveness, but to make computational suggestions visible, reviewable, and interruptible in a clinician-facing intake-support interface. To clarify scope, the evaluated engine is a local text-classification and extractive-evidence pipeline rather than a free-form proprietary LLM generator. Six datasets were analyzed: CounselingBench, Graph2Counsel, CounselBench-Eval, a CBT distortion test set, a staged CBT response-quality dataset, and AnnoMI. The primary task classified CounselingBench counselor questions and answer options into five counseling-support competencies. The best probability-producing model, TF-IDF word features with stochastic-gradient log-loss classification, achieved 0.677 accuracy and 0.532 macro-F1 on 405 held-out cases. Evidence extraction generated three evidence chips per case and reached 0.993 evidence/full-prediction agreement, indicating fidelity to the implemented classifier rather than clinical sufficiency. At a 0.70 confidence threshold combined with risk-term routing, the interface released 7.2% of cards for routine review, achieved 0.897 accuracy on released cards, and routed 97.7% of model errors to human review. External checks showed that Graph2Counsel strategy prediction achieved 0.610 micro-F1, CBT response acceptability reached 0.807 accuracy, and AnnoMI therapist-behavior classification reached 0.693 macro-F1. The findings support the card as a cautious information-architecture prototype: it can expose recommendation category, confidence, model evidence, risk flag, next-step question, and human-review action, while leaving final interpretation and clinical appropriateness to the therapist.