Paper Care Logs to Analysis-Ready Data: The ParkinSync Pipeline

ParkinSync is a serverless pipeline that takes paper-based home-care logs — the kind a family caregiver actually keeps — and turns them into a dataset you can review alongside weather and indoor-temperature context.

The current handler automatically imports normalized OCR rows into Google Sheets. An operator needs to review after import; the code does not enforce approval before writing to the ledger. Import success is not evidence that the source was transcribed correctly or that every row has been reviewed. (The review boundary is explained in Why Automatic OCR Import Still Needs Human Review.)

The architecture

• Amazon S3 — upload target for PDFs and record files
• AWS Lambda — ingestion plus scheduled jobs
• Amazon Textract — extraction of table cells from scanned logs
• AWS Secrets Manager — API keys and Google credentials
• Google Sheets API — the master ledger
• SwitchBot Open API — indoor temperature
• Visual Crossing Weather API — per-date weather context
• Python analysis scripts — schema checks and exploratory analysis over a 25-column dataset

There are two Lambdas. One ingests care logs: it fires on S3 uploads, fetches the weather for the record’s date, and appends to Google Sheets. The other logs indoor temperature: EventBridge runs it on a schedule, it reads the sensor API, and writes to the same ledger.

Why Google Sheets instead of a database

We had the usual options — DynamoDB, RDS, even SQLite on a schedule. We chose Google Sheets, and it was not a compromise.

Care logs don’t just need to be processed; they need to be trusted by non-engineers. A spreadsheet is something the caregiver themselves can open, read, correct, and export. A human-legible ledger supports post-import inspection and correction. It does not enforce that inspection or make an unchecked row trustworthy.

What’s next

The current handler already has processed-object tags and OCR-failure quarantine with a notification attempt. These are failure and reprocessing controls, not a human approval gate: tagging happens after the spreadsheet write, and tagging failure can still allow a later event to reprocess the object. Stronger review-state enforcement remains development work. A synthetic sample dataset would let others try the pipeline without household records.

ParkinSync doesn’t automate any medical decision. It is an information layer — it turns daily caregiving records into something you can verify later and bring to a care team.

The project is open source, and the product page tracks where it stands: ParkinSync · GitHub. If you are curious how this looks from the caregiver’s side, the Japanese essay 紙の介護記録は、そのままでいい is the same philosophy, minus the YAML.