pyhealth
by K-Dense-AIpyhealth helps you build clinical and healthcare deep-learning pipelines with a Dataset → Task → Model → Trainer → Metrics workflow. Use this pyhealth skill for MIMIC-III/IV, eICU, OMOP, SleepEDF, ChestXray14, EHRShot, prediction, drug recommendation, sleep staging, ICD coding, EEG events, and medical code mapping.
This skill scores 78/100 and is worth listing: it gives directory users a clear, reusable PyHealth trigger plus a concrete clinical-ML workflow model, though it is still missing some adoption aids like support files and an install command. Users should expect a solid niche skill for healthcare ML tasks, not a fully packaged toolchain.
- Strong triggerability: the description explicitly covers PyHealth, MIMIC, eICU, OMOP, EHR modeling, drug recommendation, sleep staging, and medical code mapping.
- Operational workflow is clear: the doc centers on a stable Dataset → Task → Model → Trainer → Metrics pipeline, which helps agents follow the intended execution path.
- Substantial body content: valid frontmatter, 6,391-character skill body, and multiple workflow/constraint signals suggest real instructional substance rather than a placeholder.
- No install command or companion support files are provided, so users may need to infer setup and dependencies from the prose.
- The repository is narrow and domain-specific; it is useful mainly for clinical/healthcare ML rather than general-purpose agent work.
Overview of pyhealth skill
What pyhealth is for
The pyhealth skill helps you build clinical and healthcare deep-learning workflows with PyHealth, especially when the job is to turn messy medical data into a repeatable Dataset → Task → Model → Trainer → Metrics pipeline. It is most useful for users working with EHR, physiological signals, or medical imaging who need a practical path from raw dataset to trainable experiment, not just a conceptual overview.
Who should use it
Use the pyhealth skill if you are working on MIMIC-III/IV, eICU, OMOP, EHRShot, SleepEDF, ChestXray14, or similar medical data and need help with prediction, drug recommendation, sleep staging, ICD coding, or EEG event modeling. It is a strong fit for scientific users who want a structured PyHealth guide for reproducible experimentation and code that matches the library’s abstractions.
What makes pyhealth different
The main advantage of pyhealth is its modular clinical workflow: datasets, tasks, models, trainer logic, and metrics are designed to fit together cleanly. That reduces glue code and makes it easier to swap models or tasks without rewriting the whole experiment. The tradeoff is that you need to respect the library’s pipeline shape; ad hoc prompt-generated code that skips task construction or data adapters often fails.
How to Use pyhealth skill
Install and open the right files
Install the pyhealth skill with npx skills add K-Dense-AI/claude-scientific-skills --skill pyhealth. Then open SKILL.md first, because it defines the intended workflow and the library-specific assumptions. If you need more context, read README.md, AGENTS.md, metadata.json, and any related rules/, resources/, references/, or scripts/ files in the repository.
Give the skill a complete clinical objective
A weak request like “use pyhealth for healthcare prediction” leaves too many choices open. A better prompt names the dataset, target task, data modality, and output expectation, for example: “Use pyhealth to build a readmission prediction pipeline on MIMIC-IV with structured EHR, and show the dataset/task/model/trainer setup.” If you want medical code mapping, say which code systems matter, such as ICD-10 to ATC or NDC to RxNorm.
Work in the library’s pipeline order
Start by identifying the dataset class, then define the task, then select the model, then configure the trainer, and only then compare metrics. That order matters because the skill is centered on how pyhealth composes experiments. When you ask in pipeline order, you get output that is easier to run, debug, and adapt than a generic “write me a model” prompt.
Read the repo with a decision-first lens
For pyhealth usage, the most valuable first pass is not exhaustive browsing; it is checking the skill file for supported datasets, tasks, model families, and any constraints around data preparation. Use that to decide whether your project is a good fit before investing in implementation. If your workload is outside the typical PyHealth workflow, ask for the closest supported pattern instead of forcing the library.
pyhealth skill FAQ
Is pyhealth only for clinical ML?
Yes, primarily. The pyhealth skill is meant for scientific and healthcare data work, especially structured clinical prediction and medical sequence modeling. If your task is not tied to EHR, signals, imaging, or medical codes, a generic Python or ML prompt is usually a better fit.
Do I need PyHealth already installed?
For actual implementation, yes. The pyhealth install step adds the skill instructions, but your environment still needs the PyHealth package and the datasets or preprocessing artifacts required by your project. If you are only exploring feasibility, the skill can help you judge whether pyhealth matches your use case before you commit to setup.
How is this different from a normal prompt?
A normal prompt often produces broad advice. The pyhealth skill is more useful when you want the library’s real workflow: dataset construction, task definition, model choice, training, and metrics in the expected order. That reduces the chance of getting code that looks plausible but does not align with PyHealth’s abstractions.
When should I not use pyhealth?
Do not use it if your work is not healthcare-related, if you need a general-purpose ML stack, or if your data does not fit one of the supported clinical modalities. It is also a poor fit when you want a fully custom research pipeline that ignores the dataset-task-model pattern.
How to Improve pyhealth skill
Specify the exact data shape
Better pyhealth results start with stronger inputs: dataset name, modality, target label, cohort logic, and what the model should predict. For example, “MIMIC-IV structured EHR, 30-day readmission, adult ICU cohort, binary classification” is much more actionable than “analyze patient data.” The more precise the input, the less the model has to guess about preprocessing and task framing.
State your implementation constraints
If you care about runtime, interpretability, baseline comparisons, or code simplicity, say so up front. PyHealth can support multiple model families, so your constraints should determine whether you want something like a transformer baseline, a recurrent model, or a recommendation-oriented architecture. This is especially important for pyhealth for Scientific work where reproducibility and experimental clarity matter more than novelty.
Ask for the first run, then refine
Use the first output to verify that the pipeline is structurally correct before optimizing model choice or metrics. If the result is too generic, ask the pyhealth skill to tighten one stage: dataset loading, task construction, model selection, or evaluation. Iterating one stage at a time usually produces better scientific code than requesting a full end-to-end system in one shot.
Watch for the common failure modes
The most common mistake is under-specifying the task so the output mixes incompatible dataset assumptions, label logic, or metrics. Another failure mode is requesting code without naming the source dataset, which makes pyhealth usage drift into placeholders. If you want reliable output, include a concrete task statement, a known dataset, and the metric you will use to judge success.