Timeskip

Developed by Michal Leff-Cohen, Hila Buchbut, Shira Shalit, Eran Tzarum and Itamar Shpitzer

📄 Download Project Book / Documentation

רכיבים רשמיים

קישור ל - itch

קישור לטריילר

תרשים זרימה מהבנות של ריפוי בעיסוק

Timeskip: Game-based ADHD Prediction

Timeskip is a research game for PC that collects objective behavioral signals during short play sessions and uses them to predict the likelihood of ADHD in children (around ages 10–12). It is not a therapy or “focus booster.” The goal is to unobtrusively measure gameplay features, store them, and send them to a trained model that returns an ADHD prediction.

---

Game flow (what the player does)

• Linear, single-player progression with a short, guided session: 0. Start scene (instructions): on-screen guidance shows how to move.
  • Controls shown: press Space to jump; use Left/Right arrows to move.
  • Purpose: a quick warm-up to ensure the player understands the controls.
  1. Forest exploration: collect screws to fix the time machine while avoiding hazards.
  2. Dwarf working‑memory challenge (Memory Game): short recall trials (before the river).
  3. River crossing: navigate across the river while avoiding fish and staying in bounds; reach the other side to advance.
• Sessions are short. At the end of a session, the game summarizes the player’s behavior and triggers data upload/prediction.

---

What we measure during play

The game logs fine‑grained events and aggregates them into a fixed, numeric feature vector the model expects (23 features). Examples:

• Task performance and safety
  • screws_spawned, screws_collected
  • forest_fish_collisions, forest_out_of_bounds, forest_approaches, forest_5_screws_collected
  • river_fish_collisions, river_out_of_bounds, river_win, river_approaches
• Timing and variability
  • target_rt_median, target_rt_cv (reaction‑time median and coefficient of variation for target actions)
  • dwarf_wm_rt_median, dwarf_wm_rt_cv (working‑memory response timing)
• Input dynamics
  • forest_keys, forest_key_gap_mean, forest_key_gap_cv
  • river_keys, river_key_gap_mean, river_key_gap_cv
• Working‑memory outcomes
  • dwarf_correct, dwarf_incorrect, dwarf_enters

These features are computed in a fixed order and sent exactly as a 23‑element vector to ensure compatibility with the trained classifier.

---

Scenes and measurements (per minigame)

Start scene — Controls tutorial

• Objective: teach the controls quickly (Space to jump; Left/Right arrows to move).
• Measurement: used for onboarding only; it does not contribute features to the 23‑value vector.
• Screenshot (to be added):

Forest — Exploration and collection

• Objective: collect screws to repair the time machine; avoid hazards.
• Controls: arrows to move; Space to jump.
• What we measure here (subset of the 23 features):
  • screws_spawned, screws_collected
  • forest_fish_collisions (hazard hits), forest_out_of_bounds, forest_approaches, forest_5_screws_collected
  • target_rt_median, target_rt_cv (reaction‑time to target actions like collecting screws)
  • forest_keys, forest_key_gap_mean, forest_key_gap_cv (input dynamics)
• Screenshot (to be added):

Dwarf Memory Game — Working‑memory challenge

• Objective: remember a short sequence/pattern and respond correctly.
• Controls: standard input per on‑screen prompts.
• What we measure here (subset of the 23 features):
  • dwarf_enters (number of entries/rounds)
  • dwarf_correct, dwarf_incorrect (trial outcomes)
  • dwarf_wm_rt_median, dwarf_wm_rt_cv (response timing and variability)
• Screenshot (to be added):

River — Navigation under constraints

• Objective: cross the river while avoiding fish and staying within bounds; reach the far bank.
• Controls: arrows to steer; Space to jump if applicable.
• What we measure here (subset of the 23 features):
  • river_fish_collisions, river_out_of_bounds, river_win, river_approaches
  • river_keys, river_key_gap_mean, river_key_gap_cv (input dynamics)
• Screenshot (to be added):

---

Data pipeline (how prediction works)

1. Logging during play
   • The game records clicks, key presses, outcomes, minigame events, and scene transitions.
2. Session export and Supabase
   • On session end, a compact JSON is built with per‑scene summaries plus a human‑readable “summary line”.
   • A row is posted to Supabase (table session_logs) containing: player_name, session_id, and session_json.
   • If Supabase isn’t configured, the session is saved locally instead.
3. Feature vector → model API
   • The same session is converted to a 23‑value feature vector and sent to a configured HTTP endpoint (e.g., a Hugging Face Space /predict).
   • The endpoint returns an ADHD prediction (label/probability). The game can briefly show this on‑screen.

Configuration (in Unity Inspector):

• Supabase URL, anon key, and table name
• Model API URL (and optional bearer token when using a routed endpoint)
• Toggle to enable API calls and whether to show the on‑screen prediction overlay

---

Purpose and scope

• Purpose: collect standardized gameplay features to predict ADHD likelihood using a trained model.
• Scope: this is a screening/research tool. It is not a medical diagnosis and not a focus/attention training app.

For a visual overview of the intended game flow, see the flowchart linked above.