How Minthe finds patterns
Most symptom trackers show you a percentage and leave you to trust it. We would rather show our working, because the method is the product. If you want to argue with our statistics, this page gives you everything you need to do it.
1. Your base rate comes first
Almost every wrong answer in food tracking comes from skipping this step. If you bloat on 40% of all days, then "I bloated after pasta" is not evidence about pasta. It is evidence about Tuesday.
So the first thing Minthe establishes is your personal base rate: how often you report a given outcome across all your logged days, regardless of what you ate. Every suspect is then measured as a contrast against that number, not against zero and not against some population average.
2. Exposure is presence, not portion
For each meal, a food is either present or it is not. Minthe does not attempt to model dose. That is a deliberate limitation, and it is worth being upfront about why: self-reported portion sizes are noisy enough that including them tends to add more error than signal.
Our FODMAP food list shows the group each of 349 foods carries, which is the same classification the engine uses. What this does require is identity. A flat white needs to count as a caffeine exposure and a lactose exposure, without counting the milk twice. Minthe handles this by logging the drink as one item carrying the real nutrition, plus attribution-only parts that carry no macros and exist purely so the engine sees both exposures.
3. Unlogged means unknown, never zero
If you did not log on Thursday, Thursday is unknown. It is not a symptom-free day.
This sounds pedantic and it is the difference between a useful app and a misleading one. Counting missing days as zeros invents healthy days you never recorded, and it biases results in a specific and nasty direction, because the days people fail to log are disproportionately the bad ones.
4. Delayed reactions, and four windows
Gut reactions are not instant, and different mechanisms run on different clocks. Minthe evaluates every suspect across four windows rather than assuming same-meal causation:
| Window | Typically catches |
|---|---|
| 0 to 2 hours | Fast mechanical and osmotic responses, reflux, early satiety |
| 2 to 6 hours | Fermentation in the small intestine, most classic FODMAP responses |
| Same day | Cumulative load across several meals |
| Next day | Colonic fermentation, stool form changes, slower inflammatory responses |
An app that only looks at the hour after eating will systematically miss the reactions that matter most to people with IBS-type symptoms.
5. Risk ratio, not odds ratio
For each suspect Minthe builds a two-by-two table and reports a risk ratio: how much more often you react on exposed days compared with unexposed ones.
Odds ratios are more common in research papers, and they are the wrong choice here. When an outcome is common, an odds ratio overstates the effect, sometimes dramatically. At a 30% base rate, an odds ratio can read nearly twice as large as the risk ratio computed from identical data. Since bloating is very common in exactly the people who download this app, using odds ratios would inflate every headline number in the product. A risk ratio also answers the question people actually ask: how much more often does this happen to me when I eat that.
6. Confidence, stated honestly
Every risk ratio is reported with a Wilson score interval, which behaves properly at the small sample sizes and extreme proportions that real personal data produces. The normal approximation most people reach for first breaks badly at exactly the counts a food diary generates.
Results are also expressed in natural frequencies rather than bare percentages, because "you bloated on 7 of the 9 days you ate onion, against 3 of 10 otherwise" is understood correctly by far more people than "78% vs 30%, p < 0.05". Where it helps, the app draws it as an icon array.
7. Correcting for testing hundreds of foods
This is the step almost nobody in this category does, and it is the one that most affects whether you can trust what you are shown.
If you test 200 foods at a 5% threshold, roughly 10 will look significant by pure chance. An app without a correction will confidently hand you ten innocent foods and send you on a pointless elimination diet.
Minthe applies a Benjamini-Yekutieli false discovery rate correction. The more familiar Benjamini-Hochberg procedure assumes independence or positive dependence between tests, which food data violates badly: onion and garlic co-occur constantly, bread implies wheat and yeast and often dairy. Benjamini-Yekutieli remains valid under arbitrary dependence, at the cost of being more conservative. We would rather miss a real trigger than manufacture a false one, because a false positive costs you a food group for months.
Where you have supplied real lab results, the correction is weighted, so that markers with genuine evidence behind them can raise the priority of related suspects. That weighting is bounded by a hard floor, so a prior can nudge the ranking but can never manufacture a pattern that the data does not support.
8. Four gates before anything reaches you
A suspect must pass all four. Failing any one of them keeps it invisible.
| Gate | Requirement | Why |
|---|---|---|
| Exposure count | At least 3 exposures for tentative language, 5 or more before Minthe speaks with any confidence | Below this, nothing is distinguishable from chance |
| Effect size | Risk ratio meaningfully above 1, not merely non-zero | A 4% increase is statistically detectable and practically useless |
| Interval | The confidence interval must exclude no-effect | A wide interval crossing 1 means you genuinely do not know yet |
| Contrast | Must stand apart from your personal base rate | If you react to everything equally, no single food is the story |
The visible result of all this is that Minthe says "not enough data yet" a lot, especially in the first fortnight. That is the honest answer, and we would rather give it than a confident number we cannot support.
9. What else was going on
Foods do not arrive alone, and neither do bad days. Every pattern Minthe surfaces comes with disclosure of what co-occurred, in plain language:
That is not a footnote, it is often the actual finding. Stress, sleep and menstrual cycle are treated as drivers in their own right rather than noise to be filtered out, and the engine runs against every well-logged gut symptom you track, not bloating alone.
10. Three confidence tiers, in plain words
| Tier | What Minthe is saying |
|---|---|
| Not enough data yet | Keep logging. No claim is being made in either direction. |
| Possible pattern | Something is there. Worth testing, not worth cutting out yet. |
| Consistent pattern | It has held up across enough exposures to act on, and it is still a hypothesis. |
11. Correlation is where we start, not where we stop
Everything above is observational. Observational data cannot separate onion from garlic when you almost always eat them together, and it cannot tell you whether a stressful fortnight drove both your worse eating and your worse symptoms.
This is why Minthe closes the loop with a guided with-and-without test. For a suspected trigger, that means several spaced challenge days with washout between them. For a food you think helps, it means a block design, roughly a week on and a week off, because carryover effects are real. The app schedules it around what you have already logged, and reads out the result against the same standards as everything else.
A deliberate test beats a year of passive observation, and it is the only step that turns a correlation into something you can actually rely on.
12. What this method cannot do
Stated plainly, because a method page that only lists strengths is marketing:
- It cannot separate foods you always eat together. Onion and garlic will implicate each other.
- It does not model dose, so "a little is fine, a lot is not" will not be discovered automatically.
- It depends entirely on your logging. Systematically forgetting to log bad days will skew results.
- It cannot diagnose anything. Coeliac disease, inflammatory bowel disease, SIBO and gastroparesis all need clinical testing, not a food diary.
- It will miss real triggers you eat rarely, because rare exposures cannot clear the gates.
Frequently asked
Why use a risk ratio instead of an odds ratio?
Odds ratios overstate effect size when the outcome is common, and bloating is very common in the people who use Minthe. A risk ratio answers the question people actually ask: how much more often do I react after eating this, compared with usual. For a base rate of 30%, an odds ratio can read almost twice as large as the risk ratio for the same underlying data.
What does Minthe do about testing many foods at once?
It applies a Benjamini-Yekutieli false discovery rate correction. Testing 200 foods at a 5% threshold would produce about 10 false positives by chance alone. Benjamini-Yekutieli is used rather than Benjamini-Hochberg because food exposures are strongly correlated with each other, and Benjamini-Yekutieli stays valid under arbitrary dependence.
Why does Minthe say there is not enough data instead of showing a result?
Because with three or four exposures almost nothing is distinguishable from chance. Showing a confident percentage on thin data is the single most common failure in symptom tracking apps. Minthe holds a suspect back until it clears four gates covering exposure count, effect size, interval width and contrast against your personal base rate.
Does a pattern mean a food caused my symptoms?
No. A pattern is a hypothesis, not a verdict. Foods travel together, so onion and garlic will often implicate each other, and a bad week at work can drive both worse eating and worse symptoms. That is why Minthe discloses co-occurring factors alongside every pattern and offers a guided with-and-without test to check the suspect properly.
What happens on days I forget to log?
If you did not log a symptom on a given day, Minthe treats that day as unknown, not as a symptom-free day. Counting missing days as zeros would quietly invent healthy days you never recorded and bias every result toward whatever you happened to eat when you were too unwell or too busy to log.
Minthe is a wellness and self-tracking tool, not a medical device. Patterns it surfaces are hypotheses for you to discuss with a clinician, not diagnoses. Nothing on this page is medical advice.