NR-588AI Week 1 Algorithmic Inventory: How to Write It

What NR-588AI Week 1 asks for

This course sits inside a complex system, meaning several organizations with separate budgets, regulators and record systems whose work must interlock for a patient to be safe. The concentration adds a second layer: some of the coordination between those organizations is now performed, ranked or triggered by software that learned its behavior from data rather than from a written rule. The opening stage exists to make you describe that layer accurately before you evaluate or govern anything in it.

Accurate description begins with a distinction the literature takes seriously and casual conversation does not. A deterministic rule fires when a documented value crosses a threshold, and anybody can read the rule. A statistical or machine-learned model produces a score from many inputs in a way no clinician at the bedside can reconstruct, and its behavior depends on which population it was trained on. Both are decision support. Only the second raises questions of provenance, calibration and drift, and a paper that treats a hard-coded alert and a proprietary risk model as the same object has lost the analytic thread in its first section.

The second thing to describe is influence. For each tool, say what it actually does to a decision: does it inform a clinician who may ignore it, does it order the queue in which patients are seen, does it gate access to a service, or does it act without anyone reading it. That ladder from advisory to determinative is where nearly all the ethical weight in this course lives, and it is a description task rather than an argument, which is why it belongs in week one.

The boundary question runs through everything. A model trained on the acute system's inpatients, deployed to rank referrals to post-acute providers, is making judgments about a population it never saw in training and it is doing so on behalf of an organization that did not build it and cannot inspect it. Naming who built a tool, who runs it, whose patients it was learned from and who bears the consequence of its output is the analytic spine of the entire session.

Expect a short written piece, a posted introduction to the setting and tool you will write about all term, or both. Choose a tool you can observe rather than one you have read about, because every later stage will ask for specifics about its inputs, its outputs and the workflow around it. If your section runs a discussion this week, treat it as final copy; posts do not reopen once submitted in Canvas.

The NR-588AI Week 1 method, step by step

Six moves for turning a vague sense that software is everywhere into a described inventory.

1. Reduce each rubric row to its verb before you write anything

   Identify, describe, analyze and evaluate sit at different depths, and an opening stage often mixes two of them. Write to the deepest verb present. The row that quietly costs points in a first submission is usually the support row, because an inventory feels like something you can write from memory.

2. Walk one pathway and list every automated output that touches it

   Take a single patient journey across the boundary you know best and note each place where software produced a number, a rank, a flag or a suggestion. Include the ones nobody calls artificial intelligence: acuity projections, risk scores embedded in assessments, referral matching, documentation assistance, scheduling optimization.

3. Classify each tool by mechanism, not by marketing

   Rule-based, statistical, machine-learned, or unknown because the vendor does not disclose. Unknown is a legitimate and important entry, and writing it honestly is stronger than guessing. Say what you can establish from documentation and what you cannot establish at all.

4. Record provenance for every model you can

   Whose patients it learned from, in what years, in what care setting, and how that population differs from the one it now runs on. This single column predicts most of the performance problems the rest of the course will study, and it is frequently the one piece of information a receiving organization never asked for.

5. Place each output on the influence ladder

   Advisory, prioritizing, gating or autonomous. Then say who can override it, whether the override is recorded, and whether anyone reviews the overrides. A tool nobody can override and nobody audits is the highest-risk entry in your inventory regardless of how accurate it is.

6. Close on ownership across the boundary

   For each tool, name the party that built it, the party that runs it, the party whose patients it affects and the party that answers when it is wrong. When those are four different organizations, say so plainly. That sentence sets up every governance argument you will make for the next seven weeks.

A layout and word budget for an algorithmic inventory

Below is the frame our tutors keep beside an opening inventory, sized for roughly 1,000 to 1,300 words plus a table. It is our own outline rather than anything the university issues, and your week's rubric outranks it wherever the two disagree.

Evidence craft for writing about deployed tools

Separate what the vendor claims from what you verified. Marketing material, procurement documentation, published validation studies and your own observation are four different classes of evidence with four different weights. Say which one each statement rests on. A performance figure quoted from a brochure is a claim about the vendor's confidence, not about the tool.

Cite published work on the category, not only on the product. Most deployed clinical prediction tools have a research literature behind the class of model even when the specific implementation is proprietary. Deterioration indices, readmission risk models and fall-risk scores have all been studied, and citing that literature lets you say something evidenced about a tool whose internals are closed.

Attribute frameworks and guidance with issuing body and year. Governance frameworks for health artificial intelligence, reporting standards for prediction model studies and regulatory positions on software as a medical device are all published, dated and revised. Name the body and the year inside the sentence, because in this field a two-year-old position may already have been superseded.

Every number arrives with its base, its window and its population. A model that flags twelve residents a week in a 90-bed facility is a different operational object from one that flags twelve in a 500-bed hospital. In algorithmic writing the population term is not optional, because performance figures do not travel between populations.

De-identify people and organizations alike. A regional acute care partner, a contracted electronic record vendor, a 120-bed skilled nursing facility. Naming a real product or organization adds nothing the rubric rewards and creates exposure the assignment never asked for.

Five mistakes that cost points in this week's territory

• Writing about artificial intelligence in general. A survey of what the technology could someday do is not an inventory of what is running in your pathway this month.
• Treating every alert as a model. Collapsing rule-based logic and learned models into one category removes the distinction the rest of the course depends on.
• No provenance column. A tool described without its training population cannot be evaluated for the boundary problem this concentration exists to teach.
• Enthusiasm or alarm in place of description. Both registers substitute a stance for an inventory, and neither can be scored against a description row.
• Hiding what you could not find out. Disclosure limits are findings. Writing that the vendor does not publish the training population is a stronger sentence than a plausible guess.

Before you submit

• Every tool is placed at a specific step in a described pathway
• Mechanism is stated for each tool, or explicitly marked as undisclosed
• Training population, years and setting appear wherever they can be established
• Each output sits on the influence ladder with its override route named
• Builder, operator, affected population and answerable party are identified for each tool
• All organizations, products and residents are de-identified throughout

Questions students ask about this stage

Keep going