Red Salud — 30-day readmission

A hospital network wants to know which discharges are likely to come back within 30 days, so the care-transitions team can call them. Score every encounter in data/holdout.csv.

Time: 90 minutes. Write everything in solution.py.

One row is one encounter, not one patient. Patients come back: a single patient can contribute five rows. holdout.csv contains no patient who appears in encounters.csv — the question is whether the model works on someone new.

Files

Columns

Known defects, stated as rules

1. Some encounters have discharged_on before admitted_on — the two dates were transcribed the wrong way round. Their length of stay is unknown, not negative and not its absolute value.
2. lab_hba1c is missing not at random. The test is ordered when someone suspects metabolic disease, so the blank itself carries information. Imputing the mean over it throws that away.

Tasks

T1 — length_of_stay(df) (15 pts)

Return the frame with a length_of_stay column in whole days, or a Series of the same length. Rule 1's encounters must be missing, not negative.

T2 — missingness_report(df) -> dict (15 pts)

Exactly three keys, about lab_hba1c:

T3 — expand_diagnoses(df) -> pd.DataFrame (15 pts)

Return the frame with one 0/1 column per diagnosis code, named dx_D01, dx_D02, … Each row flags exactly the codes it lists. No rows dropped.

T4 — patient_aware_split(df, valid_share) -> (train, valid) (15 pts)

Split into train and validation so that no patient_id appears in both halves, and valid is roughly valid_share of the rows (within 4 percentage points — patients have different numbers of encounters, so a quarter of patients is only about a quarter of rows).

T5 — fit_predict(train, holdout) -> np.ndarray (40 pts)

Both frames arrive raw. Return P(readmitted_30d = 1) for every holdout row, in file order. The holdout's patients are strangers to the training file; anything keyed to a patient identity will not survive.

Graded on ROC-AUC: >= 0.590 beats ranking by prior visits, >= 0.635 is competent, >= 0.665 means the diagnosis list and the missing-test indicator are earning their place.

Allowed: pandas, numpy, scikit-learn.

Visible cases — Red Salud — 30-day readmission

What CodeSignal shows next to Run tests: small inputs with the output they should produce. Run them with make drill-test. They use toy data, never the day's CSVs, so passing them is not a grade — make drill-grade is.

T1 — length_of_stay (Length of stay, including the impossible ones)

days between the dates; a negative gap is a missing stay

Input:

encounter_id admitted_on discharged_on
          E1  2025-06-01    2025-06-04
          E2  2025-06-10    2025-06-08
          E3  2025-06-05    2025-06-05

Expected:

length_of_stay == [3, NaN, 0]  (E2 is discharged before it is admitted; not -2, not 2)

T2 — missingness_report (Is the missing lab telling you something)

the readmission rate on each side of the missing lab

Input:

encounter_id  lab_hba1c  readmitted_30d
          E1        NaN               1
          E2        6.1               0
          E3        NaN               0
          E4        7.0               1
          E5        NaN               1

Expected:

{'missing': 3, 'rate_when_missing': 0.6667, 'rate_when_present': 0.5}

T3 — expand_diagnoses (A column that holds a list)

one 0/1 column per code, every row flags what it lists

Input:

encounter_id diagnosis_codes
          E1         D03;D07
          E2             D07
          E3     D01;D03;D07

Expected:

columns dx_D01 = [0, 0, 1], dx_D03 = [1, 0, 1], dx_D07 = [1, 1, 1]; 3 rows, original columns kept

T4 — patient_aware_split (Split without splitting a patient)

no patient on both sides, about a quarter of the rows

Input:

74 encounters from 40 patients (1 to 3 visits each)
valid_share = 0.25

Expected:

(train, valid): every encounter on one side; no patient_id shared; valid is roughly 25% of the rows (a quarter of patients, not of rows)

T5 — fit_predict (Predict readmission for patients you have never seen)

one probability per holdout encounter, patients never seen

Input:

toy training frame of 200 patients (every column, three swapped date pairs, half the labs missing); 31-row holdout of 20 new patients

Expected:

array of 31 floats in [0, 1], no NaN