Split Adverse Events at Severity Changes

1library(yamaa)23yamaa_spec("spec.yaml")$run()$output

1import yamaa23yamaa.spec("spec.yaml").run().output

1# R reference for the yamaa benchmark adam-adae-worsening-emergence.2# Uses the collected event records and dated findings to construct periods.3library(readr)45read_input <- function(name) {6  read_csv(paste0('/app/input/', name),7           col_types = cols(.default = col_character()),8           show_col_types = FALSE)9}10normalize <- function(x) {11  short <- !is.na(x) & nchar(x) == 16L12  x[short] <- paste0(x[short], ':00')13  x14}15ae <- read_input('ae.csv')16fa <- read_input('fa.csv')17adsl <- read_input('adsl.csv')18fa$FADTC <- normalize(fa$FADTC)19ae <- ae[order(ae$STUDYID, ae$USUBJID, as.integer(ae$AESEQ)), ]20ranks <- c(MILD = 1L, MODERATE = 2L, SEVERE = 3L)21output <- list()22previous_study <- previous_subject <- NA_character_23aseq <- 0L24for (i in seq_len(nrow(ae))) {25  event <- ae[i, ]26  subject <- event$USUBJID27  study <- event$STUDYID28  if (!identical(study, previous_study) ||29      !identical(subject, previous_subject)) aseq <- 0L30  previous_study <- study31  previous_subject <- subject32  trt <- normalize(adsl$TRTSDTM[33    adsl$STUDYID == study & adsl$USUBJID == subject34  ])[1]35  onset <- normalize(event$AESTDTC)36  ending <- normalize(event$AEENDTC)37  findings <- fa[fa$STUDYID == study & fa$USUBJID == subject &38                   fa$FALNKID == event$AELNKID, ]39  times <- sort(unique(findings$FADTC))40  sevs <- rep(NA_character_, length(times))41  grades <- rep(NA_integer_, length(times))42  sev <- NA_character_43  grade <- NA_integer_44  for (j in seq_along(times)) {45    updates <- findings[findings$FADTC == times[j], ]46    if (anyDuplicated(updates$FATESTCD)) stop('Duplicate finding at event/time')47    new_sev <- updates$FASTRESC[updates$FATESTCD == 'SEV']48    new_grade <- updates$FASTRESC[updates$FATESTCD == 'TOXGR']49    if (length(new_sev) && !is.na(new_sev)) sev <- new_sev50    if (length(new_grade) && !is.na(new_grade)) grade <- as.integer(new_grade)51    sevs[j] <- sev52    grades[j] <- grade53  }54  before <- if (is.na(trt)) integer() else which(times < trt)55  baseline_sev <- baseline_grade <- NA_integer_56  if (length(before)) {57    last <- tail(before, 1)58    baseline_sev <- unname(ranks[sevs[last]])59    baseline_grade <- grades[last]60  }61  emergent <- FALSE62  for (j in seq_along(times)) {63    if (!is.na(onset) && !is.na(trt)) {64      if (onset >= trt) emergent <- TRUE65      else if (times[j] >= trt) {66        rank <- unname(ranks[sevs[j]])67        sev_worse <- !is.na(rank) && !is.na(baseline_sev) && rank > baseline_sev68        grade_worse <- !is.na(grades[j]) && !is.na(baseline_grade) &&69          grades[j] > baseline_grade70        emergent <- emergent || sev_worse || grade_worse71      }72    }73    aseq <- aseq + 1L74    output[[length(output) + 1L]] <- data.frame(75      STUDYID = study, USUBJID = subject, ASEQ = aseq,76      AESEQ = as.integer(event$AESEQ), AEDECOD = event$AEDECOD,77      AESTDTC = onset, AEENDTC = ending, ASTDTM = times[j],78      AENDTM = if (j < length(times)) times[j + 1L] else ending,79      TRTSDTM = trt, AESEV = event$AESEV, AETOXGR = event$AETOXGR,80      ASEV = sevs[j], ATOXGRN = grades[j],81      TRTEMFL = if (emergent) 'Y' else NA_character_82    )83  }84}85dir.create('/app/output', showWarnings = FALSE)86write_csv(do.call(rbind, output), '/app/output/adae.csv', na = '')

1# Reference solution for the yamaa benchmark adam-adae-worsening-emergence (Python track).2#3# Written from the benchmark's full prompt and its inputs alone, as an agent4# would write it. Harbor's oracle agent runs it to check that the prompt can5# be solved and that the grader scores a correct answer 1.67"""Build ADAE with one record per subject per adverse event per analysis8period, carrying forward the most recently collected severity and toxicity9grade, and flagging treatment-emergent worsening after first exposure."""1011import csv12from datetime import datetime1314INPUT = "/app/input"15OUTPUT = "/app/output"1617SEV_RANK = {"MILD": 1, "MODERATE": 2, "SEVERE": 3}181920def read_csv(path):21    with open(path, newline="") as f:22        return list(csv.DictReader(f))232425def parse_dt(text):26    text = (text or "").strip()27    if not text:28        return None29    return datetime.fromisoformat(text)303132def fmt_dt(value):33    if value is None:34        return ""35    return value.strftime("%Y-%m-%dT%H:%M:%S")363738def main():39    adsl = read_csv(f"{INPUT}/adsl.csv")40    ae = read_csv(f"{INPUT}/ae.csv")41    fa = read_csv(f"{INPUT}/fa.csv")4243    # First exposure per subject (blank for untreated subjects).44    trtsdtm = {}45    for row in adsl:46        trtsdtm[row["USUBJID"]] = row["TRTSDTM"]4748    # FA findings per (subject, link id): lists of (time, value) per test.49    findings = {}50    for row in fa:51        key = (row["USUBJID"], row["FALNKID"])52        entry = findings.setdefault(key, {"SEV": [], "TOXGR": [], "times": set()})53        when = parse_dt(row["FADTC"])54        entry["times"].add(when)55        if row["FATESTCD"] == "SEV":56            entry["SEV"].append((when, row["FASTRESC"]))57        elif row["FATESTCD"] == "TOXGR":58            entry["TOXGR"].append((when, row["FASTRESC"]))59    for entry in findings.values():60        entry["SEV"].sort(key=lambda item: item[0])61        entry["TOXGR"].sort(key=lambda item: item[0])6263    def most_recent(items, start):64        """Value most recently collected at or before start, else None."""65        value = None66        for when, val in items:67            if when <= start:68                value = val69            else:70                break71        return value7273    out_rows = []74    # Events in AESEQ order within each subject, subjects as collected.75    events = sorted(76        ae,77        key=lambda r: (r["USUBJID"], int(r["AESEQ"])),78    )79    subjects = []80    for row in events:81        if row["USUBJID"] not in subjects:82            subjects.append(row["USUBJID"])83    events.sort(key=lambda r: (subjects.index(r["USUBJID"]), int(r["AESEQ"])))8485    for subj in subjects:86        subj_events = [r for r in events if r["USUBJID"] == subj]87        exposure = parse_dt(trtsdtm.get(subj, ""))88        aseq = 089        for ev in subj_events:90            onset = parse_dt(ev["AESTDTC"])91            entry = findings.get((subj, ev["AELNKID"]), {"SEV": [], "TOXGR": [], "times": set()})92            starts = sorted({onset} | entry["times"])9394            periods = []95            for i, start in enumerate(starts):96                if i + 1 < len(starts):97                    end = fmt_dt(starts[i + 1])98                else:99                    # Last period ends at the event end; an ongoing event100                    # (no AEENDTC) has no AENDTM.101                    end = ev["AEENDTC"] if ev["AEENDTC"] else ""102                asev = most_recent(entry["SEV"], start)103                grade = most_recent(entry["TOXGR"], start)104                periods.append(105                    {106                        "start": start,107                        "ASTDTM": fmt_dt(start),108                        "AENDTM": end,109                        "ASEV": asev if asev else "",110                        "ATOXGRN": int(grade) if grade not in (None, "") else None,111                    }112                )113114            # Treatment-emergent flag.115            flags = [""] * len(periods)116            if exposure is not None:117                if onset >= exposure:118                    flags = ["Y"] * len(periods)119                else:120                    # Reference: the last pre-treatment period's values.121                    pre = [p for p in periods if p["start"] < exposure]122                    ref_sev = SEV_RANK.get(pre[-1]["ASEV"]) if pre else None123                    ref_grade = pre[-1]["ATOXGRN"] if pre else None124                    emergent = False125                    for i, p in enumerate(periods):126                        if p["start"] < exposure:127                            continue128                        if not emergent:129                            sev_rank = SEV_RANK.get(p["ASEV"])130                            sev_up = (131                                sev_rank is not None132                                and ref_sev is not None133                                and sev_rank > ref_sev134                            )135                            grade_up = (136                                p["ATOXGRN"] is not None137                                and ref_grade is not None138                                and p["ATOXGRN"] > ref_grade139                            )140                            if sev_up or grade_up:141                                emergent = True142                        if emergent:143                            flags[i] = "Y"144145            for p, flag in zip(periods, flags):146                aseq += 1147                out_rows.append(148                    {149                        "STUDYID": ev["STUDYID"],150                        "USUBJID": subj,151                        "ASEQ": aseq,152                        "AESEQ": ev["AESEQ"],153                        "AEDECOD": ev["AEDECOD"],154                        "AESTDTC": ev["AESTDTC"],155                        "AEENDTC": ev["AEENDTC"],156                        "ASTDTM": p["ASTDTM"],157                        "AENDTM": p["AENDTM"],158                        "TRTSDTM": trtsdtm.get(subj, ""),159                        "AESEV": ev["AESEV"],160                        "AETOXGR": ev["AETOXGR"],161                        "ASEV": p["ASEV"],162                        "ATOXGRN": "" if p["ATOXGRN"] is None else p["ATOXGRN"],163                        "TRTEMFL": flag,164                    }165                )166167    columns = [168        "STUDYID", "USUBJID", "ASEQ", "AESEQ", "AEDECOD", "AESTDTC",169        "AEENDTC", "ASTDTM", "AENDTM", "TRTSDTM", "AESEV", "AETOXGR",170        "ASEV", "ATOXGRN", "TRTEMFL",171    ]172    with open(f"{OUTPUT}/adae.csv", "w", newline="") as f:173        writer = csv.DictWriter(f, fieldnames=columns)174        writer.writeheader()175        writer.writerows(out_rows)176177178if __name__ == "__main__":179    main()

1/* opensas reference for yamaa's adam-adae-worsening-emergence benchmark.2   Construct periods from one AE and its collected, dated FA findings. */3data ae;4  length STUDYID USUBJID AELNKID AEDECOD AESTDTC AEENDTC AESEV AETOXGR $1024;5  infile '/app/input/ae.csv' dsd dlm=',' firstobs=2 truncover lrecl=32767;6  input STUDYID : $char1024. USUBJID : $char1024. AESEQ7    AELNKID : $char1024. AEDECOD : $char1024. AESTDTC : $char1024.8    AEENDTC : $char1024. AESEV : $char1024. AETOXGR : $char1024.;9  if lengthn(AESTDTC)=16 then AESTDTC=cats(AESTDTC,':00');10  if lengthn(AEENDTC)=16 then AEENDTC=cats(AEENDTC,':00');11run;12data fa;13  length STUDYID USUBJID FALNKID FATESTCD FAOBJ FASTRESC FADTC $1024;14  infile '/app/input/fa.csv' dsd dlm=',' firstobs=2 truncover lrecl=32767;15  input STUDYID : $char1024. USUBJID : $char1024. FASEQ16    FALNKID : $char1024. FATESTCD : $char1024. FAOBJ : $char1024.17    FASTRESC : $char1024. FADTC : $char1024.;18  if lengthn(FADTC)=16 then FADTC=cats(FADTC,':00');19  SEVRANK=.;20  if FATESTCD='SEV' then do;21    select(FASTRESC);22      when('MILD') SEVRANK=1; when('MODERATE') SEVRANK=2;23      when('SEVERE') SEVRANK=3; otherwise SEVRANK=.;24    end;25  end;26run;27data adsl;28  length STUDYID USUBJID TRTSDTM $1024;29  infile '/app/input/adsl.csv' dsd dlm=',' firstobs=2 truncover lrecl=32767;30  input STUDYID : $char1024. USUBJID : $char1024. TRTSDTM : $char1024.;31  if lengthn(TRTSDTM)=16 then TRTSDTM=cats(TRTSDTM,':00');32run;33proc sql;34  create table changes as35    select STUDYID,USUBJID,FALNKID,FADTC as ASTDTM,36      max(SEVRANK) as SEVCOLN,37      max(case when FATESTCD='TOXGR' then input(FASTRESC,best32.)38        else . end) as GRDCOL39    from fa group by STUDYID,USUBJID,FALNKID,FADTC;40  create table joined as41    select a.*,c.ASTDTM,c.SEVCOLN,c.GRDCOL,b.TRTSDTM42    from changes as c inner join ae as a43      on c.STUDYID=a.STUDYID and c.USUBJID=a.USUBJID44        and c.FALNKID=a.AELNKID45    left join adsl as b46      on a.STUDYID=b.STUDYID and a.USUBJID=b.USUBJID;47quit;48proc sort data=joined;49  by STUDYID USUBJID AESEQ ASTDTM;50run;51data states;52  length ASEV $1024;53  set joined;54  by STUDYID USUBJID AESEQ;55  retain ASEVN ATOXGRN;56  if first.AESEQ then do; ASEVN=.; ATOXGRN=.; end;57  if not missing(SEVCOLN) then ASEVN=SEVCOLN;58  if not missing(GRDCOL) then ATOXGRN=GRDCOL;59  select(ASEVN);60    when(1) ASEV='MILD'; when(2) ASEV='MODERATE';61    when(3) ASEV='SEVERE'; otherwise ASEV='';62  end;63run;64/* Read the next start by scanning periods in reverse chronological order. */65proc sort data=states;66  by STUDYID USUBJID AESEQ descending ASTDTM;67run;68data periods;69  length AENDTM NEXTSTART $1024;70  set states;71  by STUDYID USUBJID AESEQ;72  retain NEXTSTART;73  if first.AESEQ then NEXTSTART='';74  if not missing(NEXTSTART) then AENDTM=NEXTSTART;75  else AENDTM=AEENDTC;76  NEXTSTART=ASTDTM;77run;78proc sql;79  create table pretreatment as80    select STUDYID,USUBJID,AESEQ,max(ASTDTM) as BASEDTM81    from periods where not missing(TRTSDTM) and ASTDTM<TRTSDTM82    group by STUDYID,USUBJID,AESEQ;83  create table baseline as84    select a.STUDYID,a.USUBJID,a.AESEQ,85      a.ASEVN as PRESEV,a.ATOXGRN as PREGRD86    from periods as a inner join pretreatment as b87      on a.STUDYID=b.STUDYID and a.USUBJID=b.USUBJID88        and a.AESEQ=b.AESEQ and a.ASTDTM=b.BASEDTM;89  create table compared as90    select a.*,b.PRESEV,b.PREGRD91    from periods as a left join baseline as b92      on a.STUDYID=b.STUDYID and a.USUBJID=b.USUBJID and a.AESEQ=b.AESEQ;93  create table worsening as94    select STUDYID,USUBJID,AESEQ,min(ASTDTM) as FIRSTWRS95    from compared where not missing(TRTSDTM) and ASTDTM>=TRTSDTM and96      ((not missing(ASEVN) and not missing(PRESEV) and ASEVN>PRESEV) or97       (not missing(ATOXGRN) and not missing(PREGRD) and ATOXGRN>PREGRD))98    group by STUDYID,USUBJID,AESEQ;99  create table flagged as100    select a.*,b.FIRSTWRS from periods as a left join worsening as b101      on a.STUDYID=b.STUDYID and a.USUBJID=b.USUBJID and a.AESEQ=b.AESEQ;102quit;103proc sort data=flagged;104  by STUDYID USUBJID AESEQ ASTDTM;105run;106data result;107  length TRTEMFL $1;108  set flagged;109  by STUDYID USUBJID;110  retain ASEQ;111  if first.USUBJID then ASEQ=0;112  ASEQ=ASEQ+1;113  TRTEMFL='';114  if not missing(AESTDTC) and not missing(TRTSDTM) then do;115    if AESTDTC>=TRTSDTM or116      (not missing(FIRSTWRS) and ASTDTM>=FIRSTWRS) then TRTEMFL='Y';117  end;118run;119proc sql;120  create table final as121    select STUDYID,USUBJID,ASEQ,AESEQ,AEDECOD,AESTDTC,AEENDTC,122      ASTDTM,AENDTM,TRTSDTM,AESEV,AETOXGR,ASEV,ATOXGRN,TRTEMFL from result;123quit;124proc export data=final outfile='/app/output/adae.csv' dbms=csv replace; run;