Kalpa Health - fde-academy-lab/c2-content-factory GitHub Wiki
Kalpa Health
Diagnostic labs and clinic operations across India and South-East Asia. The unit where a wrong answer has a consequence that is not symmetric, and where the correct behaviour is sometimes to refuse.
Code
HLT. The home of refusal rules, late labels, and the difference between an output that is faithful and an output that is right.
Where it sits in the locked ladder
Build 1, Week 3. The stakeholder is Dr Priya Menon, COO, and she has Meera's problem in a different business: test volumes grew 5 percent against a plan of 18, and she does not know which branch of her business is short. Her data is real, messy and unfamiliar: lab bookings, sample collections, results, invoices and appointment no-shows across six Indian cities.
Five sub-problems, three groups each, and every one of them is the Week 1 and Week 2 method in a domain the room has never seen:
| # | The sub-problem | The Week 1 or 2 move it transfers |
|---|---|---|
| 1 | Where does lab revenue come from and which branch is short | The revenue tree, for a diagnostics business |
| 2 | Bookings fell in two cities in Q2 | The investigation ladder on booking data |
| 3 | Invoices and collections disagree | Reconciliation between the billing export and the payment feed |
| 4 | The no-show rate looks worse in one clinic | Real or noise, and what a fair comparison needs |
| 5 | A free-home-collection campaign "lifted bookings 9 percent" | Cause or coincidence |
What changes in the transfer is the vocabulary (a booking is an order, a test is an item, a no-show is a cancellation) and the cost of an error, since a missed diagnosis is not a missed sale. A group that maps Health onto the Week 1 method in its own words has done the transfer; a group that waits to be told the mapping has not.
From Week 5 onward Health is also the second worked example on teaching days once Financial Services has had its turn, usually on no-shows and claims.
1. Who runs it, and what each of them wants
| Role | Measured on | What they want to hear | What they will resist |
|---|---|---|---|
| Operations lead, diagnostic labs | Turnaround time and rejection rate | That this month is the best on record | A maturity cutoff that makes the current month unreportable |
| Clinic operations director | Slot utilisation and no-show rate | That empty slots can be predicted away | Arithmetic showing the calls cost more than the slots save |
| Clinical governance lead | Amendments, incidents, complaints | That nothing is being missed | Throughput arguments that trade accuracy for volume |
| Head of Claims | Claim acceptance rate and days to payment | That rejections are falling | A cohort view that shows the current month is simply incomplete |
| Head of Patient Support | Contacts handled, satisfaction | That an assistant can take the load | A refusal set that makes the assistant look less capable |
The tension that generates most cards here: clinical caution and operational throughput are both legitimate, and they point in opposite directions. This is the unit where a learner has to say a number is not the whole story to somebody whose job is the number.
2. The data estate
| What exists | The catch it carries |
|---|---|
| Sample registration and test completion timestamps | Rejections are recorded on average nine days after registration, and the lag differs by reason |
| Rejection reasons where present | No field says whether a sample is still in progress, so incomplete and clean look identical |
| Appointment records, booking time, clinic, slot | Operational fields such as slot_released are populated after the outcome and leak the target |
| Test panels and line items | Panels were unbundled into line items, so test counts and order counts moved in opposite directions |
| Report text, and amendments to reports | The amendment rate is the honest quality metric and nobody wants to own it |
| Claim submissions and adjudications | Right-censored in exactly the same way as rejections, one layer further out |
The absence that defines the unit: nothing records why a patient did not arrive. Every no-show model is built on a target whose cause is unobserved, and every intervention changes the target it was trained on.
3. The KPIs, and how each one is gamed
| KPI | What people say it is | The denominator that decides it | How it breaks or gets gamed |
|---|---|---|---|
| Sample rejection rate | Rejected divided by registered | Cohorts of samples with a full maturity window | Every month looks clean while it is running, and every closed month keeps rising for weeks |
| Turnaround time | Registration to report | Per test type, with send-outs separated, at a percentile | Exclude send-out tests and the number is excellent and meaningless |
| No-show rate | Missed appointments divided by booked | Booked, with reschedules counted once | A reschedule counted as both a no-show and a new booking double counts the denominator |
| Slot utilisation | Slots used divided by slots offered | Slots that could have been filled | Offer fewer slots and utilisation rises |
| Repeat test rate | Same test repeated within a window | Repeats that were not clinically intended | A repeat can be a quality failure or good medicine, and the data cannot tell you which |
| Report amendment rate | Reports corrected after issue | All issued reports | The one metric here that cannot be gamed upward, which is why it is the one to watch |
| Refusal correctness | Whether an assistant declined appropriately | A seeded set of questions it should refuse and a set it should not | Over-refusal is invisible to every grounding metric, and it is how staff stop using a tool |
The KPI this unit exists to teach: a rate reported only on settled cohorts. Once a learner internalises that a running period is not comparable to a closed one, half the analytics traps in the bank become visible on sight.
4. Case families, mapped to modules
| Module | Case family here | The teaching version |
|---|---|---|
| 1 Foundations of AI and Data | Rejection rate, turnaround time, no-show rate, and the maturity cutoff | S-HLT-AN-03, the month that always looks clean |
| 2 Applied Machine Learning | No-show prediction, escalation models, cost asymmetry between the two errors | S-HLT-ML-02, where the strongest feature is populated after the outcome |
| 3 Deep Learning and Neural Networks | Structured clinical sequences over a patient's history | Where a point-in-time feature set stops being optional |
| 4 Natural Language Processing | Claims document processing and referral note extraction | Short, high-consequence text with heavy abbreviation |
| 5 and 6 Generative AI | Report summarisation with strict refusal rules, patient support assistants | S-HLT-GA-03, and the checks that need no answer key |
| 7 Production AI | Version pinning, frozen regression sets, inter-rater agreement as a noise floor | An evaluation you can re-run on demand |
| 8 and 9 Agentic AI | An assistant that must escalate rather than answer, and how that boundary is enforced | Where "the prompt says not to" stops being a control |
5. The situations hosted here
| Page | Cards |
|---|---|
| Analytics | S-HLT-AN-03 (the month that always looks clean), S-HLT-AN-16, S-HLT-AN-17 |
| Machine learning | S-HLT-ML-02 (the feature that already knew), S-HLT-ML-14, S-HLT-ML-15 |
| GenAI and agents | S-HLT-GA-03 (we changed nothing and it got worse), S-HLT-GA-17 |
6. What Health must never be used for
- Clinical claims. No card asserts what a test means, what a result implies, or what a clinician should do. Every card is about the operational system around the medicine.
- Any real patient data or anything resembling it. The repository is public and this is the unit where that rule matters most.
- A regulatory assertion. Cards may say a governance lead would ask for an audit trail. No card states what any actual regulation requires unless somebody has verified it and dated it.
- A first modelling lesson. Consequence asymmetry needs a room that already trusts the trainer.