(CASE STUDY · 07)

MINDSCOPE

Data Cleaning
Machine Learning
Streamlit
UX | UI
Docker

MindScope is a mental wellbeing self-check built on 2 standard questionnaires, PHQ-9 for depression and GAD-7 for anxiety. A scikit-learn model reads both scores with 10 more answers, including age, sleep, stress and work, and returns a Low, Moderate or High risk band with advice to match.

ROLE: Data, ML & UI
TIMELINE: 5 Months
YEAR: 2026
TEAM: Solo

(MY ROLE)

Cleaned and profiled a synthetic mental health dataset in a notebook.
Trained 3 classifiers on 12 features and shipped the best one, a logistic regression, with its encoders.
Designed and built the app twice: one long form in February, a 4-step wizard in June.
Wrote the Dockerfile, with a build step that puts the share-card tags where crawlers can read them.

(WHY IT EXISTS): PHQ-9 and GAD-7 are short and standard, so a self-check built on them takes minutes. The goal was to take the analysis out of the notebook and into an app.
(THE FIRST VERSION DID TOO MUCH AT ONCE): Version 1 put all 16 questions and the lifestyle fields on one long page. On the results page, the advice sat folded inside expanders.
(WHAT CHANGED): In June the form became a 4-step wizard with softer answer labels and a sage and clay palette. Version 1 stopped with an encoding error when gender was Other or Prefer not to say. Options the encoders never saw now fall back to a known class.

Process

01
Data & Scales

Cleaned the Kaggle Global Mental Health Dataset 2025 in a notebook: 2,500 records, 15 columns, 10 countries. Scores map to the standard PHQ-9 and GAD-7 severity bands.
02
Label & Model

The risk label is the mean of the two scores: under 5 Low, under 12 Moderate, otherwise High. Encoded 7 categorical columns and trained 3 classifiers on a stratified 80/20 split, keeping the best test score.
03
First Release

Wrapped the saved model and encoders in a Streamlit app that loads them once per process. The app and a Dockerfile were committed on the same day in February.
04
Redesign

Rebuilt the interface in June as a 4-step wizard that keeps each answer across steps. Phone layouts came the same day.

How It Works

Answer 16 questions over 4 steps. The app adds up the PHQ-9 and GAD-7 scores, encodes the other answers and asks a saved model for a risk band. Every answer stays in the session. Nothing is written down.

The Check-In

The subject is sensitive, so the goal was reflection, not alarm. Warm sand, sage and clay, serif headings, and answers in plain words: Not really, Here & there, Quite a lot, Nearly always. The chips shade from sage to clay as an answer grows.

MindScope home page in a browser: the MindScope header with its PHQ-9 · GAD-7 · AI chip, the Home, Assessment and About buttons, the sage hero reading Understand your mind, gently and clearly, Begin Assessment and Learn More, and the tops of the 3 feature cards.

Step 3 of 4, Anxiety · GAD-7, on its own: the progress bar and the first 4 questions on graded chips, answered Quite a lot, Quite a lot, Here & there and Quite a lot.

In The Detail

A running score sits under each questionnaire and moves with every answer. The result leads with the band and the model's confidence, then the scores. Below, the same question in version 1 and version 2.

PHQ-9 questions 3 and 4 mid-answer, Here & there and Quite a lot, with the running score card below: Your PHQ-9 so far, 7 / 27, Mild.

The risk banner, Assessed risk · 98% model confidence, Moderate Risk, Take Action, above the PHQ-9 12/27 and GAD-7 10/21 metric cards and the top of the PHQ-9 gauge.

The same questions in both versions. February, version 1: a dark Mental Health Screening Test page with Basic Information, then PHQ-9 question 1 as a numbered list from 0, Not at all, to 3, Nearly every day. June, version 2, laid over it: step 2 of 4 with questions 1 to 3 on graded chips.

(SYSTEM 01 · RISK MODEL)

Risk Model

A logistic regression, kept from 3 classifiers and cached with st.cache_resource. It reads 12 numbers and returns Low, Moderate or High, with a probability for each.

(NOT ONE RECORD IS A REAL PERSON): The Kaggle dataset is generated, not collected from patients. The model has never seen a real person's answers.
(THE LABEL IS A FORMULA OF 2 INPUTS): Both scores that define the risk label are also among the 12 features. So the model mostly relearns the formula, and its test score reflects that.

(SCOPE)

questions across 2 scales: 16
features in every prediction: 12
records in the synthetic dataset: 2,500

The sample run: the 12 numbers predict_risk builds from the answers, the encoders and their fallback for unseen options, the logistic regression's probabilities with Moderate at 0.985, and get_risk_level beside the PHQ-9, GAD-7 and risk bands.

(SYSTEM 02 · GUIDANCE)

Guidance

Every band opens its own advice: 50 tips in 14 categories, 12 for Low, 17 for Moderate and 21 for High. Each band has its own colour and banner. A High result opens with Immediate Actions, then treatment, crisis lines, daily wellness and self-care; a Low result keeps to activity, sleep, mental wellness and lifestyle. Every result ends with a reminder that this is not a diagnosis.