Preparing your stability data
Paste or drop the table you already have. It does not need to be cleaned first: units, condition formats, <LOQ results and drifting RRTs are handled automatically, and every decision is listed in the data log.
The usual layout
One row per sample, one column per result:
| Batch | Time (days) | Condition | RRT 0.45 | RRT 0.82 | Assay (%LC) |
|---|---|---|---|---|---|
| A | Initial | Initial | <0.05 | 0.02 | 100.2 |
| A | 7 | 50°C/75%RH | 0.08 | 0.04 | 99.9 |
| A | 2 | 70°C/75%RH | 0.14 | 0.07 | 99.4 |
The values above are made up to show the format.
One row per result (LIMS exports)
Most LIMS export one row per result instead. Paste that as it is; it is recognised and reshaped automatically:
| Batch | Condition | Time point | Test | Analyte | Result | Units |
|---|---|---|---|---|---|---|
| A | 50°C/75%RH | Day 7 | Related substances | RRT 0.45 | 0.08 | % area |
| A | 50°C/75%RH | Day 7 | Related substances | RRT 0.82 | 0.04 | % area |
| A | 50°C/75%RH | Day 7 | Assay | Assay | 99.9 | % LC |
| A | 50°C/75%RH | Day 7 | Water (KF) | Water | 2.1 | % w/w |
- Rows for assay are recognised by the test or analyte name, and rows for other tests (water, dissolution, appearance) are left out.
- Summary rows such as "Largest unspecified impurity" are left out, because each peak is modelled on its own. "Total impurities" rows are used as the total.
- A peak not listed for a sample where other peaks were reported counts as below the reporting threshold, just like a blank cell.
- Replicates (the same peak twice for one batch, time point and condition) are kept as separate results.
- Time points such as "T0", "Day 7", "Week 2" or "3 months" are understood.
Everything set aside is listed in the data log, so you can check nothing was missed.
Columns that are recognised
| Column | What works |
|---|---|
| Time | Numbers with the unit in the header: hours, days, weeks, months or years ("Time (weeks)", "Days"), or in the cell: "3M", "M3", "3 months", "Week 2", "T3". "Initial", "T0" and "0" mark the starting point. |
| Condition | One column such as 50°C/75%RH, 50/75 or 50C 75% RH, or separate temperature and humidity columns. Kelvin and humidity as a fraction (0.75) are converted, tolerances (25 ± 2 °C) are ignored, and a range such as 2–8 °C is read as its midpoint, 5 °C. |
| Degradants | One column per peak, headed by RRT or name ("RRT 0.82", "Imp B"). Footnote marks (0.21*) and comments in brackets (0.12 (OOS)) are ignored. Text that is not a result, such as "NR" or "see note", is read as no result and listed in the data log. Peaks within ±0.02 RRT that never appear in the same sample are treated as one degradant. |
| Assay | Optional. Used to model assay loss from the initial value against the lower limit. |
| Total degradants | Optional. If missing, the total is the sum of the reported peaks, leaving out peaks seen only under harsh stress. |
| Statistics and partial totals | Columns such as "Assay RSD", "Total unspecified" and "Largest unknown" are left out, so they are never modelled as results. |
| One row per result | Instead of one column per peak: a peak name (or RRT) column and a result column, optionally with test name, batch and units. See above. |
Workbooks laid out for people
Titles and product details above the table, a heading such as "Impurities (%)" over a second header row of peak names, units in a row of their own, and a condition written once and merged down its block of rows are all read. Cells formatted as percentages in Excel are read as shown (0.12%, not 0.0012). Blank cells filled from the row above are listed in the data log, and row numbers in the log match the rows in Excel.
Results below the limit
<LOQ, ND, <0.05, ND (<0.05), BLOQ and blank cells in samples where other peaks were reported are all accepted. By default they are treated as known to be below the limit (statistically censored), which avoids the bias of replacing them with zero or half the limit. You can choose either of those instead.
Study design that predicts well
- At least three temperatures, and two or more humidity levels so the humidity effect can be estimated.
- Pull points that take the fastest-growing degradant a good way towards its limit at the harsher conditions. If nothing grows by more than 10% of the way to the limit, predictions are long extrapolations and you will see a warning.
- An initial (time zero) result for every peak, even if it is <LOQ.
The tool checks these for every degradant and says when the design limits what can be concluded.
Excel workbooks
.xlsx files are read in your browser. If the workbook has several sheets with data, you choose which one to use. Old .xls files need saving as .xlsx or CSV first, or you can copy the table and paste it.
Checking the prediction with real-time data
Include long-term or intermediate results (for example 25 °C/60 %RH at 3, 6 and 12 months) in the same table. On the settings page they are ticked under Real-time data, which leaves them out of the fit and uses them only to check the prediction. This happens automatically for conditions at 30 °C and below when three or more hotter temperatures are left to fit. Each result is compared with the 95% prediction interval for a single new measurement, which covers both the uncertainty in the fit and the scatter between measurements. The results page, the report and the Results JSON show which results fall within it. Expect about one result in twenty to fall outside by chance; several outside, or a trend, means the accelerated model does not describe long-term storage well. Untick a condition to fit it with the rest.
Adding later results
When new pulls come in (the 6- or 12-month real-time results, say), reopen the saved analysis and choose Add results on the results page. Paste the new rows with their header row, or drop a file. Columns are matched by heading, so they can be in a different order. A peak seen for the first time becomes a new column, read as below the reporting threshold in the earlier samples. Rows already in the table are skipped, so you can drop the whole updated table. Existing rows keep their numbers, so excluded points and locked models still apply. The data log and the report record when the results were added and from where. Save the analysis again afterwards.
Seeing the data used
Open Data used on the results page to see the table the models were fitted to: merged RRTs as one column, blanks filled in as below the reporting threshold, and the total and assay loss as modelled. Each result is coloured by its use: fitted, possible outlier, real-time check (within or outside the prediction), excluded by you, or in a response that was not modelled. The settings page shows every row of your table, with the rows held out for the real-time check highlighted. The data used can also be downloaded as CSV.
Saving and reopening an analysis
Save results › Analysis file downloads a .json file holding your data, column choices, settings, limits, locked models and excluded points. To carry on later, open the calculator and choose Open saved analysis, or drop the file on the first page. The file carries a fingerprint of the data, so a file whose data was edited is refused. It is not the same as the Results JSON below, which is for other software and can't be reopened.
Results as data (JSON)
Save results › Results JSON downloads every result in a structured file you can load into a database, a notebook or your own models. For each degradant, the total and the assay it holds:
- the chosen model, why it was chosen, and whether you or the tool chose it;
- the fitted kinetics with standard errors and the covariance matrix: activation energy (kJ/mol), humidity sensitivity B (per %RH), the rate at 50 °C/50 %RH and, where used, the time exponent and plateau;
- every candidate model with its fit statistics;
- shelf life, 95% lower bound and probability of staying within the limit at each storage condition;
- the cleaned data points, any you excluded with the reason, and the data log.
The file also carries the settings, the data fingerprint, the engine version and the study details you entered (product, active ingredient, dosage form, packaging, batch). Shelf lives beyond 10 years are written as null with beyondHorizon: true. Field names stay the same within a schemaVersion, so files from different studies can be combined.