Smart Bathroom Health Monitoring: Toilet Bowl Image Classification Case Study
How Shaip delivered a structured medical image annotation pipeline classifying 10,000 toilet bowl images per batch with 5 primary event classes, Bristol Stool Scale clinical scoring, and 20+ attribute tags — powering non-invasive digital health monitoring AI.
Project Overview
As non-invasive digital health monitoring moved into smart bathroom deployment, the client needed a clinically grounded annotation pipeline capable of classifying high volumes of telecentric camera imagery against a structured decision tree with attribute-level granularity.
Shaip built the end-to-end annotation pipeline covering image usability assessment, primary event classification, Bristol Stool Scale scoring, attribute tagging, and batch-cycle QA handoff — producing clinically applicable datasets for gastrointestinal health monitoring AI.
Key Status
Images per Batch
10,000
Delivery Cycle
14 days
Primary Event Classes
5
Bristol Scale Range
1–7
Challenges
- Building a clinically grounded classification schema spanning 5 primary event classes plus 20+ attribute tags
- Operating on a strict 14-day batch cycle with 10,000 images delivered per cycle
- Applying Bristol Stool Scale 1–7 scoring consistently across all stool events
- Handling monochromatic, obstructed, and artifact-affected imagery with usability-first triage
- Coordinating client QA handoff with structured day-7 and day-14 milestones
Solution
Decision Tree Classification
Shaip designed a structured decision-tree workflow covering five primary event classes — Empty, Urine, Stool, Flushing, and Toilet Soap — with combinable tags for compound events (e.g., flushing combined with stool). An Event Unclear tag handled ambiguous cases while preserving identifiable event combinations.
Clinical Bristol Scale Scoring
For all stool events, annotators assigned a Bristol Stool Scale score from 1 to 7, with Unclear or Mix flags available for difficult cases. This clinical scoring layer made the dataset directly applicable to gastrointestinal health monitoring AI training.
Rich Attribute Tagging
Beyond primary classification, each image received attribute tags covering urination stream visibility, toilet paper presence, blood detection (including suspected blood, blood clots, and blood crumbs), corrosion markings, camera dirt, image obstruction, flash artifacts, uncentered framing, low volume, floating content, and toilet leaks. Monochromatic images in orange or purple-blue tones received additional color flags.
Image Usability Triage
Every image was first assessed for usability and relevance. Pixelated, overexposed, green, gray, non-telecentric, or off-topic images were flagged as ignore. This triage layer prevented compromised inputs from entering the labeled dataset, protecting model training quality downstream.
Two-Week Batch Cycle & QA Handoff
Shaip operated on a strict two-week batch cycle with clearly defined milestones. After Shaip completed tagging, the client's internal QA team reviewed half the batch by day 7 and the full batch by day 14. Rejected images were flagged and returned for correction, ensuring continuous quality improvement across batches.
Project Scope
| Dataset Type | Volume per Batch | Event Classes | Clinical Scoring | Attribute Tags | Cycle |
|---|---|---|---|---|---|
| Smart toilet image classification | 10,000 images | 5 primary + combinations | Bristol Stool Scale 1–7 | 20+ tags | 14 days |
Outcomes
- Established a clinically grounded annotation pipeline for non-invasive digital health monitoring AI
- Standardized 5-class primary event classification with combinable tag logic
- Delivered Bristol Stool Scale 1–7 scoring across every stool event
- Built 20+ attribute tags spanning clinical conditions, image quality, and visual artifacts
- Maintained strict 14-day batch cycles with structured client QA handoff
Overall, Shaip helped transform a high-volume medical image classification requirement into a clinically structured, production-ready annotation pipeline — one capable of supporting non-invasive digital health monitoring, gastrointestinal AI, and at-home clinical decision support applications at scale.
Shaip understood that this wasn’t routine image tagging — it was clinical annotation. Their Bristol Scale execution, attribute tag rigor, and batch cycle discipline gave us training data we could confidently take into model deployment.
– Head of AI, Digital Health Platform
