Koji Miyazawa

Data scientist.

At a medical data company, I analyse electronic health records and cancer genomic panel test results from hospitals, for pharmaceutical companies. For example, how often lung side effects appear with anticancer drugs, how often gene variants appear, and finding side effects in free-text clinical notes.

I also work on digital transformation (DX) and AI adoption for small and mid-sized businesses. I build business systems such as an ERP made from scratch, an on-premise AI server, and software for accounting and education. I recently released GogaQ, an iPhone and iPad app for learning languages from real sentences.

Other areas I work in: claims data, doctor surveys (conjoint analysis), Bayesian statistics and causal inference. I publish the tools I make here.

My open-source projects

Project	What it does
jpomop	Converts Japanese DPC inpatient data to the OMOP Common Data Model v5.4. Built from the public specification and synthetic data only. On PyPI: `pip install jpomop`.
partworth	Hierarchical Bayes choice-based conjoint analysis in Python (PyMC), with Sawtooth-style outputs and a market simulator. On PyPI: `pip install partworth`.
natural-japanese-writing	An agent skill for Claude Code, Codex and Cursor that rewrites AI-sounding Japanese into natural Japanese without changing the facts.

Open-source contributions

Synthea International #23, merged: a Japan configuration for Synthea, the synthetic patient data generator. It adds Japanese regions, prefectures and municipalities, hospitals and clinics, income data, and patient cost sharing.
awesome-japanese-nlp-resources #33, merged: added natural-japanese-writing to the list of Japanese NLP resources.
ArviZ (arviz-stats) #489, merged: hdi now returns NaN for both bounds when the input contains NaN and skipna is False.
PyHealth #1310, merged: added the OMOP tables observation_period, visit_detail, observation and device_exposure to the default OMOP config, with test data.
admiral #3206, merged: compute_duration() now errors when a datetime has no time zone or the two time zones differ. admiral is the pharmaverse R package for building clinical trial analysis datasets. Listed in the README acknowledgments.
NumPyro #2381, merged: with validation on, TransformedDistribution now raises when the base support is outside the domain of the first transform. Before, Normal(0, 1) with ExpTransform().inv built without error and gave NaN samples.
PyMC #8524, fixed in #8522 with me as co-author: StudentT and SkewStudentT logcdf returned NaN at ±inf, so one-sided truncated draws were all NaN.
PyMC-Marketing #3250, merged: the NestedLogit inclusive value dropped the max utility, so with λ=1 nest probabilities did not reduce to the multinomial logit. Added the missing term and a closed-form test.
Orca #28921, merged: sidebar text that stayed in English under the Japanese UI now goes through translation, and 25 Japanese strings with the wrong meaning are fixed. Credited in the v1.4.243 release notes.
OHDSI CommonDataModel #819, merged: typo fixes in the OMOP CDM documentation.

Worked examples in Bayesian statistics

Executed Jupyter notebooks, each on one method and one public dataset.

bayesian-pairwise-ranking: Bradley-Terry, Davidson and Plackett-Luce in PyMC, applied to LLM arena votes, food preferences and tennis
bayesian-ab-decision-kit: sequential A/B testing with expected-loss stopping rules
bayesian-psychometrics: Rasch, 2PL and graded response models
dp-customer-segmentation: Dirichlet process mixtures on retail RFM features
uplift-modeling-benchmark: S/T/X meta-learners on the Criteo uplift dataset
bayesian-volatility-regimes: stochastic volatility and regime-switching HMMs
backtest-overfitting-check: CSCV/PBO, Deflated Sharpe Ratio and MinBTL

kintone plugins

Ten open-source plugins for kintone, written in TypeScript. Each README is available in English and Japanese.

	Plugin	What it does
	command-palette	Ctrl+K command palette: jump to apps, views and records from the keyboard
	record-diff-viewer	Compare two records side by side with field-level and inline text diff
	status-time-tracker	Track how long records spend in each process status and locate bottlenecks
	round-robin-assigner	Assign new records to team members in rotation or by current workload
	draft-autosave	Auto-save form input and restore it after a lost tab or session timeout
	markdown-copy	Copy records and list views to the clipboard as Markdown tables, lists or TSV
	image-annotator	Draw arrows, shapes and text on attached images, keeping the original file
	image-compressor	Compress and resize images in the browser before attaching, stripping EXIF location data
	dummy-data-generator	Generate realistic Japanese test data, reproducible with a fixed seed
	theme-switcher	Dark and dim themes with per-user switching and OS auto mode