5-year-old Female — Type 1 Diabetes Mellitus, Insulin Dosing Volatility & Recurrent Hypo/Hyperglycemia, West Bengal

What the mix teaches (educator note)

This case is an unusually clear demonstration of a hybrid human–AI care team: a lay caregiver generating dense, high-quality primary data; a rotating panel of remote physicians providing periodic clinical steer; and multiple large language models providing on-demand pattern analysis of uneven reliability — including at least one instance of a fabricated literature citation, later caught and corrected by a participant. The stakeholder mix illustrates both the promise (24/7 responsive analysis, translation, citation-seeking) and the necessary guardrails (human verification, correction of AI over-claiming) of AI-augmented remote chronic disease management.

You are doing something genuinely hard extremely well: logging almost every meal, dose, and reading for over two years is rare and valuable. The big swings you're seeing between very low and very high sugars aren't a sign you're failing — they're mostly coming from two fixable habits: (1) snacks like milk, biscuits, and fruit going in without any insulin, which lets sugar climb quietly before the next meal's dose, and (2) giving a big "correction" shot when sugar is very high, which can push it too low a few hours later and start the cycle again. The single change most likely to help is learning to count carbs for every snack, not just meals — even a small dose for the morning milk or the evening banana. It's also worth asking your care team about a continuous sugar sensor (CGM), even for a short trial period, since it would show you the overnight picture you can't currently see with finger-pricks. And please keep the sick-day rule in mind: even if she's not eating, she still needs her long-acting insulin — never skip it during a fever or vomiting.

Two-year longitudinal capillary log (finger-prick only, no CGM) shows extreme glycemic variability (range 39–595 mg/dL) that the remote group's pattern analysis attributes chiefly to unaddressed prandial dosing technique rather than basal insufficiency: repeated basal (Tresiba) dose increases made in response to post-hypoglycemic rebound highs have each been followed within 24–72h by a new hypoglycemic trough, consistent with over-basalization rather than a genuine rising basal requirement. Working insulin sensitivity factor used by the family (1 U : 70 mg/dL) appears to systematically overestimate correction needs; several isolated correction events in the log suggest an empiric ISF closer to 1:50–60. No CGM data exists to confirm nocturnal patterns (Somogyi vs. dawn phenomenon) — recommend CGM trial (even 1 sensor / 2–3 months, diagnostic use) if cost-prohibitive for continuous wear. One fever/vomiting episode with glucose 512–595 mg/dL was managed at home without ketone testing (no meter available) — worth confirming this did not represent unrecognized mild DKA and ensuring the family has ketone strips and a written sick-day protocol. No confirmed HbA1c value is documented in the log for calendar year 2025; recommend confirming interval labs are being drawn, not just estimated GMI from self-monitoring.

Adherence to logging is excellent and should be preserved as-is — do not add reporting burden. The main coordination gap is between the frequent remote group discussion and comparatively infrequent in-person visits with the local doctor; ensure any dose-strategy changes discussed remotely (e.g. carb-counting approach, basal timing) are explicitly relayed to and confirmed by the treating physician before being acted on, to avoid the family receiving conflicting instructions from multiple sources. CGM affordability is a recurring, unresolved logistics item — worth actively researching state-level (West Bengal T1DM integrated care model) or NGO/UNICEF-linked support schemes rather than leaving this as an open question across months. Multiple different AI assistants have been used inconsistently for analysis within the same thread, with at least one fabricated citation identified — recommend standardizing on a single, clearly-labeled analytic tool with a "verify before acting" norm attached to every AI-generated recommendation.

• Pattern-recognition skill exercised: distinguishing a true basal insulin deficiency from a bolus-driven Somogyi rebound — the case shows this being mistaken repeatedly, with basal dose changed in the wrong direction because the "input" (a bolus/snack timing error) and "output" (fasting glucose) were separated by many hours.
• Easy to miss on first read: the log's own retrospective realization that "uncovered nuts and milk" had likely always caused post-breakfast glucose in the 300–400 range, but this was never checked with a timed reading until very late in the case — a reminder that absence of a measurement is not absence of an effect.

Socratic questions for discussion:

• If a patient's fasting glucose is high the morning after a very high pre-bed reading, how would you distinguish "the basal dose is too low" from "the basal dose is too high and rebounding"? What single additional data point would settle it fastest?
• Given the same insulin regimen produced wildly different outcomes depending only on snack coverage, what does that suggest about where scarce educational resources (a single course, a single clinic visit) should be spent first in a resource-limited setting?
• Several AI-generated analyses in this case were clinically sound, but at least one fabricated a supporting citation when asked to justify a claim. What workflow would let a care team benefit from AI pattern-spotting while reliably catching that kind of error?