ECAA-workflow/config/project-class-keywords.yaml at main · SuLab/ECAA-workflow · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
# Parallel to config/modality-keywords.yaml but decides
# the ProjectClass enum. Conservative defaults: Bioinformatics wins ties
# (D7). Non-bio hits must be unambiguous (frozen SAP vocabulary, forecast
# vocabulary).
#
# Match semantics follow config/modality-keywords.yaml: a keyword is a
# whole-word match when alphanumeric-only, or a substring match when it
# contains whitespace or punctuation.

classes:
  - id: clinical_trial
    keywords:
      - "statistical analysis plan"
      - sap
      - saps
      # Bare "prespecified" / "pre-specified" / "pre specified" removed
      # False-positives on bio-domain prose like
      # "QC fails a prespecified floor", "prespecified gene panel",
      # "prespecified contrast" trip every paper-recreation prompt
      # into clinical_trial routing. The v4 composer then prefers
      # the clinical_trial_analysis archetype (CDISC ADaM inputs),
      # port-match fails on FASTQ-vs-CDISC mismatch, throws
      # GoalUnreachable, and emits a 0-task DAG. The narrower
      # multi-word forms below preserve real clinical-trial-SAP
      # intake. Mirrors the bare-"arms" / bare-"endpoint" removals
      # already documented below.
      - "prespecified analysis plan"
      - "prespecified statistical analysis plan"
      - "prespecified sap"
      - "prespecified primary endpoint"
      - "prespecified secondary endpoint"
      - "prespecified clinical endpoint"
      - "prespecified inclusion criteria"
      - "prespecified exclusion criteria"
      - "pre-specified analysis plan"
      - "pre-specified statistical analysis plan"
      - "pre-specified sap"
      - "pre-specified primary endpoint"
      - "pre-specified secondary endpoint"
      # Bare "arms" is excluded: it false-positives on bio-domain
      # prose (e.g., "albuterol and dex+albuterol arms across four
      # donors" in bulk-RNA-seq intake gets misrouted to
      # clinical_trial). The multi-word clinical-trial keywords
      # below cover real SAP intake without tripping on generic
      # experimental-design prose; narrower "treatment arm" /
      # "trial arm" tokens can be added if the catalog needs them.
      # Bare "endpoint" / "endpoints" are excluded too — they
      # false-positive on "CNGB HTTPS endpoint", "REST API
      # endpoints", "dataset endpoint URL", etc. The multi-word
      # forms below cover real clinical-trial intake and don't
      # trip acquisition-stage URL vocabulary.
      #
      # Bare "primary endpoint" / "secondary endpoint" are
      # excluded: they false-positive on bio-domain statistical
      # prose like "AUROC is not a valid primary endpoint for a
      # generalization claim". The narrower trial-context forms
      # below cover real SAP intake. Co-occurring clinical signals
      # (sap, itt, per-protocol, cdisc, adam, randomized controlled
      # trial, rct, hazard/odds ratio) still classify real
      # clinical-trial prose unambiguously.
      - "co-primary endpoint"
      - "co-primary endpoints"
      - "trial primary endpoint"
      - "trial secondary endpoint"
      - "study primary endpoint"
      - "study secondary endpoint"
      - "protocol primary endpoint"
      - "protocol secondary endpoint"
      - "endpoint analysis"
      - "endpoint adjudication"
      - "intent to treat"
      - "intent-to-treat"
      - itt
      - "per-protocol"
      - "per protocol"
      - cdisc
      - adam
      - sdtm
      - "hazard ratio"
      - "odds ratio"
      - "randomized controlled trial"
      - rct
      # Bare "phase N" / "phase i/ii/iii" patterns are excluded:
      # they false-positive on genomics references like
      # "1000 Genomes" reference-panel descriptors (GWAS
      # paper-recreation prose). The narrower trial-context forms
      # below cover real clinical-trial intake via either the
      # multi-word forms here or co-occurring `rct` /
      # `randomized controlled trial` tokens.
      - "phase iii trial"
      - "phase iii clinical trial"
      - "phase iii randomized"
      - "phase iii study"
      - "phase iii rct"
      - "phase 3 trial"
      - "phase 3 clinical trial"
      - "phase 3 randomized"
      - "phase 3 study"
      - "phase 3 rct"
      - "phase ii trial"
      - "phase ii clinical trial"
      - "phase ii randomized"
      - "phase ii study"
      - "phase ii rct"
      - "phase 2 trial"
      - "phase 2 clinical trial"
      - "phase 2 randomized"
      - "phase 2 study"
      - "phase 2 rct"

  - id: time_series_forecast
    keywords:
      - "time series"
      - "time-series"
      - forecast
      - forecasting
      - arima
      - arma
      - sarima
      - seasonality
      - seasonal
      - stationarity
      - autocorrelation
      - "rolling window"
      - "forecast horizon"
      - rmse
      - mape
      - "ljung-box"
      - "state space"
      - "structural break"