Central data monitoring report

Coordinating centre view - Synthetic Adaptive Platform Trial in Critical Illness

2550 participants · 3066 randomisations · 25 sites in 5 countries · rule set 455d0fba76e0 · generated 28 August 2026

Sites needing attention

Ordered by entry-delay drift, then by critical findings. Deliberately not a single composite score: a score hides which of the three problems a site has, and the coordinating centre needs to know which call to make.

Site Name Country Critical findings Endpoint records Not evaluable % incomplete Drift (days/month)
SE-02 Västerlid Sjukhus SE 21 86 3 23.3 1.69
FI-02 Lahdenpää Yliopistollinen Sairaala FI 12 75 2 29.3 0.18
SE-03 Almstrand Lasarett SE 4 46 1 23.9 0.15
NL-01 Sint-Aldegonde Ziekenhuis NL 14 133 1 24.8 0.10
SE-01 Björkhagen Universitetssjukhus SE 16 123 3 20.3 0.08
NL-02 Maasstad Ziekenhuis Noord NL 11 93 6 34.4 0.08
NL-03 Duinrand Medisch Centrum NL 6 69 1 29.0 0.07
DK-02 Vestbro Sygehus DK 13 154 4 26.0 0.06
FI-01 Koivuranta Keskussairaala FI 12 104 4 30.8 0.03
DK-05 Havnestad Sygehus DK 8 131 2 22.9 0.03

Entry-delay drift

No validation rule catches drift. Every record at a drifting site is perfectly valid - it was simply entered late, which is not a rule violation. Only the trend across months shows a site falling further behind, and only measuring it against each site’s own initiation shows that it is not merely a young site still settling in.

Findings across the trial

Rule Name Severity Form Findings Participants Sites
4 RNG-002 heart_rate_plausible critical daily_icu 106 99 25
5 RNG-003 temperature_plausible critical daily_icu 88 87 25
2 LOG-002 not_alive_after_death critical daily_icu 42 18 14
16 XDM-003 vital_status_consistent_where_windows_agree critical outcome_30d 25 12 9
12 TMP-001 discharge_not_before_admission critical outcome_30d 24 24 16
13 TMP-004 ae_onset_before_randomisation critical adverse_events 23 23 16
11 STR-004 single_randomisation_per_domain critical randomisation 22 11 13
10 STR-003 participant_id_unique_to_one_site critical screening 10 5 9
14 XDM-001 death_date_consistent_across_domains critical outcome_30d 7 3 6
6 STR-001 required_field_present major daily_icu 1682 916 25
9 STR-001 required_field_present major screening 644 375 25
1 LOG-001 no_gap_in_daily_records major daily_icu 138 137 24
8 STR-001 required_field_present major randomisation 136 116 24
7 STR-001 required_field_present major outcome_30d 100 98 22
3 RNG-001 weight_plausible major screening 38 28 12
15 XDM-002 icu_admission_consistent_across_domains major outcome_30d 12 5 9

How well do the rules actually work?

The validation engine is scored against a catalogue of defects deliberately injected into the data, so its performance is measured rather than asserted.

Defect Name Expected rule Injected Detected Recall
D01 missing_required_field STR-001 2635 2559 97.1%
D02 missing_daily_record_gap LOG-001 46 45 97.8%
D03 out_of_range_value RNG-001 16 16 100.0%
D03 out_of_range_value RNG-002 103 103 100.0%
D03 out_of_range_value RNG-003 88 87 98.9%
D04 impossible_date_sequence TMP-001 25 24 96.0%
D05 ae_before_randomisation TMP-004 23 23 100.0%
D06 duplicate_participant_id STR-003 6 5 83.3%
D07 double_randomisation_same_domain STR-004 8 8 100.0%
D08 inconsistent_vital_status LOG-002 14 14 100.0%
D09 late_entry_drift none - monitoring signal 739 0 n/a
D10 terminal_digit_preference none - monitoring signal 453 0 n/a
D11 unit_conversion_failure RNG-001 31 23 74.2%
D12 ae_under_reporting none - monitoring signal 17 0 n/a
D13 allocation_update_not_applied none - monitoring signal 196 0 n/a

Overall: 97.1% - 2907 of 2995 defect records that a rule targets.

A further 1405 records belong to defect types that no rule in this milestone targets. They are reported as n/a rather than as 0%, because scoring a rule set against defects it was never written to catch would misrepresent it, and omitting them would misrepresent the defect catalogue. These are the statistical signals - entry drift, terminal-digit preference, adverse-event under-reporting - and the drift section above is the first of them to be addressed.

Findings not traceable to an injected defect

Not the same as false positives. A finding with no matching injected defect is usually a genuine problem the simulation created incidentally, or a real ambiguity in the data model.

Rule Name Severity Findings Traceable to injection Not traceable
LOG-001 no_gap_in_daily_records major 138 46 92
XDM-003 vital_status_consistent_where_windows_agree critical 25 0 25
XDM-002 icu_admission_consistent_across_domains major 12 0 12
LOG-002 not_alive_after_death critical 42 35 7
XDM-001 death_date_consistent_across_domains critical 7 0 7
STR-004 single_randomisation_per_domain critical 22 16 6
STR-001 required_field_present major 2562 2560 2
RNG-003 temperature_plausible critical 88 87 1
RNG-001 weight_plausible major 38 38 0
RNG-002 heart_rate_plausible critical 106 106 0
STR-003 participant_id_unique_to_one_site critical 10 10 0
TMP-001 discharge_not_before_admission critical 24 24 0
TMP-004 ae_onset_before_randomisation critical 23 23 0

LOG-001 is the clearest example. No form in this trial carries an ICU discharge date, so a gap in daily records caused by discharge and readmission is indistinguishable from a gap caused by missing data entry. Querying both is what a data manager would actually do, and the fix is a schema change rather than a rule change.

What the ingest layer had to undo

Every transformation applied between the site’s export and the analysis dataset is logged. This table is the evidence of what was done to the data.

Transformation Sites Log entries Values affected
2 date_format_normalised 20 220 46918
1 blank_to_na 25 510 6757
4 decimal_separator_normalised 5 15 3673
3 datetime_format_normalised 20 20 2589
5 encoding_transcoded 1 5 1341
6 unit_converted 2 2 270

Primary endpoint completeness by site

A day with no daily record counts as unknown, never as a day free of life support. Missing records therefore understate the endpoint. The percentage incomplete is a direct measure of how much each site’s data is degrading the outcome the trial exists to measure.

n not_evaluable n_complete n_incomplete died_within_window mean_all mean_complete_only median_all total_unknown_days
3058 99 2233 726 960 16.18 14.55 22 1608

The mean over all records and the mean over complete records differ. That difference is not a rounding detail: it is larger than the one-day margin the analysis plan uses to declare two arms practically equivalent, which is why the endpoint reports its own completeness rather than returning a single number.

All data in this report is synthetic, generated by the code in this repository from a fixed seed. Site names are invented. This is an independent educational project with no affiliation to any real trial, hospital or research group.