A machine learning study of 2,500 synthetic teen profiles, exploring what actually predicts depression risk — with a model you can try yourself.
Adjust the inputs below and watch how the model's estimate changes. This is a toy built on made-up data — it says nothing about any real person, including you.
This is not a screening or diagnostic tool, and not a substitute for talking to a professional about mental health.
"Lowkey stressed." "Lowkey tired." Quiet words for things that aren't quiet at all. LOWKEY takes them seriously: a machine learning study of what actually signals teen depression risk, paired with honest evidence and clear paths to help.
Move the conversation beyond screen-time blame, toward what the evidence shows matters most — and make asking for help feel as normal as asking "kamusta ka?"
A generation where no young person struggles quietly, and no adult around them is afraid to ask.
Each record is one adolescent: habits, sleep, activity, and self-reported stress and anxiety. The target: a three-class depression-risk label (low / medium / high).
Low risk is the majority class (50.6%) — one reason macro-averaged F1 was used, so every class counts equally.
First five records from the raw survey table.
| Age | Media h/day | Platform | Sleep h | Stress | Anxiety | Risk |
|---|---|---|---|---|---|---|
| 19 | 3.8 | 6.5 | 6 | 6 | medium | |
| 16 | 6.5 | Other | 5.8 | 9 | 7 | high |
| 17 | 2.9 | Other | 7.0 | 3 | 2 | low |
| 19 | 7.9 | TikTok | 5.0 | 10 | 8 | medium |
| 15 | 1.2 | TikTok | 8.8 | 1 | 2 | low |
Three model families compared on mean testing macro F1 across stratified splits: Logistic Regression, Random Forest, and Gradient Boosting.
Testing macro F1 vs regularization strength C (log scale). L1 at C = 0.1 wins the baseline round.
Mean testing scores per model. Gradient Boosting takes every metric.
C = 0.1
300 trees · max depth 5
200 estimators · learning rate 0.01
Data. 2,500 synthetic adolescent records, 11 predictors, target depression_risk ∈ {low, medium, high} (1,266 / 650 / 584 — imbalanced, hence macro-averaged metrics that weight every class equally). Categorical variables integer-encoded; all features standardized with StandardScaler fit on training data only.
Evaluation protocol. 75/25 train–test splits, stratified on the target; every hyperparameter configuration evaluated over 10 trials with varying random seeds, selecting by mean testing macro F1. Final models refit with random_state=42.
Search spaces. Logistic Regression (liblinear, one-vs-rest): penalty ∈ {L1, L2} × C ∈ {0.001, 0.01, 0.1, 1, 10, 100} → best L1, C=0.1 (F1 0.666). Random Forest: trees ∈ {100, 200, 300} × depth ∈ {None, 5, 10} → best 300 trees, depth 5 (F1 0.759). Gradient Boosting: estimators ∈ {100, 200} × learning rate ∈ {0.01, 0.1, 0.2} → best 200, lr=0.01 (F1 0.774, accuracy 0.799).
Convergent feature importance. Gradient Boosting's impurity-based importances and L1's sparsity agree: L1 zeroed out 5 of 11 features and retained stress (mean |coef| 2.05) and sleep (0.24) — an implicit feature-selection check that the 93% stress finding isn't an artifact of one model family.
Browser demo disclosure. The interactive tool above runs the L1 model's published per-class coefficients in JavaScript, with feature standardization statistics estimated from the data's plausible ranges and intercepts calibrated to reproduce the training class balance — a faithful qualitative approximation, not the pickled model. Exact parity would require exporting the fitted sklearn pipeline (e.g., to ONNX). Stated limitations: synthetic data, correlational design, self-reported measures, and no external validation cohort.
Read this first: one feature explaining 93% of a model's predictions is not typical in real behavioral research — it's the signature of a synthetic dataset whose labels likely weighted stress heavily. Treat 93% as a demo of how feature importance works, not evidence about real teens. Stress and sleep as top factors is consistent with real clinical research (see the peer-reviewed studies below) — but this exact number is a property of this dataset, not a measurement of the real world.
Both top models agree: self-reported stress level drove 93% of Gradient Boosting's feature importance, with sleep a distant second. Social media time barely registered.
Share of total importance (log-scaled axis so the small contributors stay visible).
Three takeaways from 2,500 profiles and three model families — and what they suggest for the adults and teens reading this.
Stress level alone accounted for 93% of the model's predictive information. For parents and teachers, the practical takeaway is about where attention goes: rather than starting with "how long were you online?", start with "what's been weighing on you lately?" — and make space to listen to the answer.
Sleep duration ranked a consistent second across models — and it's the input that tips borderline cases in the tool above. Adolescents need roughly 8–10 hours of sleep; protecting a regular sleep routine is one of the most concrete, evidence-aligned things a household can support.
Social-media hours, platform choice, and screen use before bed together contributed less than 2% of the signal. Heavy screen use may be how distress looks from the outside — a way of coping or withdrawing — so restricting devices alone can address the visible behavior while leaving the underlying need unmet.
How to read these results responsibly
Synthetic data means association isn't causation — this shows what occurs together, not what causes what. Low grades don't cause depression; they're usually a consequence of distress, not its source. A falling grade is a reason to ask "kamusta ka?" — never to add pressure. No model output can assess a real person. Depression is diagnosable and treatable — concerns about mood, sleep, or withdrawal are best brought to a physician, psychologist, or counselor.
One peer-reviewed study per factor — chosen for scale and rigor, not for making the strongest headline:
Notice how this pattern echoes the study's own finding: the two factors with the deepest, most consistent evidence base — stress and sleep — are also the two that carried this model's signal, while academic performance and social media, despite real associations in the literature, contributed comparatively little here.
The actual PHQ-9 questionnaire, adapted for teens by the GLAD-PC team and hosted by the American Academy of Child & Adolescent Psychiatry. Unlike the demo above, this is a real instrument used in real clinics — meant to be filled out, then reviewed with a clinician, same as here.
Before you start: this screens, it doesn't diagnose — only a professional can do that. Nothing you answer is saved or sent anywhere — it disappears when you close the page, and it's never used to train this site's model. Show your results to a parent, doctor, or school counselor — that's the most useful next step.
In the past year, have you felt depressed or sad most days, even if you felt okay sometimes?
How difficult have these problems made it for you to do your work, take care of things at home, or get along with other people?
Has there been a time in the past month when you have had serious thoughts about ending your life?
Have you EVER, in your WHOLE LIFE, tried to kill yourself or made a suicide attempt?
Source: Modified with permission by the GLAD-PC team from the PHQ-9 (Spitzer, Williams & Kroenke, 1999), Revised PHQ-A (Johnson, 2002), and the CDS (DISC Development Group, 2000). Guidelines for Adolescent Depression in Primary Care, Version 2, 2010. Hosted by AACAP: aacap.org (PDF).
Beyond the everyday habits covered in the model above, here's how to support a friend in the moment, and where to find real help in the Philippines. (Looking for a screening tool? Try the PHQ-9 Screening section.)
Most teens tell a friend before an adult — so you might be the first to know. Needing help is normal, and so is asking for it. You don't need perfect words:
To open the door: "Napansin ko lang, parang iba ka these days. Kamusta ka?" · "You don't have to explain everything. I just want you to know I noticed."
To keep it open: "That sounds really heavy. I'm glad you told me." · "Gusto mo bang samahan kita to talk to the guidance counselor? We can go together."
Two things to remember: listening is not the same as fixing — you don't have to solve it. And if a friend talks about hurting themselves, that's never a secret to keep, even if they ask you to. Telling a trusted adult or calling a hotline together is how you stay a good friend, not how you betray one.
When concerns go beyond what a check-in can hold, these professional societies maintain directories of qualified specialists across the Philippines:
Free, confidential, and staffed by trained responders. Save this number in your phone.
More avenues of support: your school's guidance office, your barangay health center, and the National Center for Mental Health (ncmh.gov.ph). Under the Philippine Mental Health Act (RA 11036), access to mental health services is a legal right. Number verified as of July 2026.
LOWKEY follows the Philippine Data Privacy Act of 2012 (RA 10173) and NPC guidance — mainly by collecting almost nothing to begin with. The safest data is the data you never collect.
Since no personal data is processed on a server, LOWKEY doesn't need NPC registration as a Personal Information Controller — but it still follows RA 10173's core principles: transparency, legitimate purpose, and proportionality (collect nothing beyond what's needed, which here is nothing at all).
Even with no personal data held here, RA 10173 guarantees you the right to be informed, access your data, object, correct or erase it, request portability, and file a complaint with the National Privacy Commission. If a future version ever collects data for research, it would only be with your explicit, separate consent.
LOWKEY is a class project, not a finished product. Here's what's missing or approximate, honestly stated.
Have a suggestion?
This project improves through feedback. If something here felt wrong, unclear, or missing — especially from teens, parents, or mental health professionals — that input is exactly what the next version needs.
Sending opens your email app with everything pre-filled, addressed to the project team — nothing is sent from this page directly, and no account is needed. This is for feedback about the project only; for urgent concerns, use the hotline above instead.
THE REAL CALL TO ACTION
Sometimes the most helpful response is asking, gently and without judgment,
“kamusta ka?”
Check in on the young people in your life. Ask about stress, listen more than you advise, and protect their sleep — the data, and clinical experience, agree that both matter.
If you or someone you know is struggling, help is available and reaching out is a sign of strength — talk to a physician, mental health professional, school counselor, or a trusted adult, or call the hotlines listed above, including the 24/7 NCMH Crisis Hotline at 1553.