Use hpstat for yli.turnbull to enable computation of confidence intervals

This commit is contained in:
RunasSudo 2023-10-20 21:11:56 +11:00
parent 675422246f
commit 14c4054a47
Signed by: RunasSudo
GPG Key ID: 7234E476BF21C61A
2 changed files with 100 additions and 42 deletions

View File

@ -45,7 +45,7 @@ The mandatory dependencies of this library are:
Optional dependencies are:
* [hpstat](https://yingtongli.me/git/hpstat), for *IntervalCensoredCox*
* [hpstat](https://yingtongli.me/git/hpstat), for *turnbull* and *IntervalCensoredCox*
* [matplotlib](https://matplotlib.org/) and [seaborn](https://seaborn.pydata.org/), for plotting functions
* [mpmath](https://mpmath.org/), for *beta_ratio* and *beta_oddsratio*
* [PyCryptodome](https://www.pycryptodome.org/), for *pickle_write_encrypted* and *pickle_read_encrypted*
@ -72,7 +72,7 @@ Relevant statistical functions are all directly available from the top-level *yl
* Survival analysis:
* *kaplanmeier*: Kaplan–Meier plot
* *logrank*: Log-rank test
* *turnbull*: Turnbull estimator plot for interval-censored data
* *turnbull*: Turnbull estimator plot, including pointwise confidence intervals, for interval-censored data
* Input/output:
* *pickle_write_compressed*, *pickle_read_compressed*: Pickle a pandas DataFrame and compress using LZMA
* *pickle_write_encrypted*, *pickle_read_encrypted*: Pickle a pandas DataFrame, compress using LZMA, and encrypt

View File

@ -15,9 +15,14 @@
# along with this program. If not, see <https://www.gnu.org/licenses/>.
import numpy as np
import pandas as pd
from scipy import stats
import statsmodels.api as sm
import io
import json
import subprocess
from .config import config
from .sig_tests import ChiSquaredResult
from .utils import Estimate, check_nan
@ -130,25 +135,27 @@ def calc_survfunc_kaplanmeier(time, status, ci, transform_x=None, transform_y=No
return xpoints, ypoints, None, None
def turnbull(df, time_left, time_right, by=None, *, step_loc=0.5, transform_x=None, transform_y=None, nan_policy='warn'):
def turnbull(df, time_left, time_right, by=None, *, ci=True, step_loc=0.5, transform_x=None, transform_y=None, nan_policy='warn', fig=None, ax=None):
"""
Generate a Turnbull estimator plot, which extends the KaplanMeier estimator to interval-censored observations
The intervals are assumed to be half-open intervals, (*left*, *right*]. *right* == *np.inf* implies the event was right-censored. Unlike :func:`yli.kaplanmeier`, times must be given as numeric dtypes and not as pandas timedelta.
The intervals are assumed to be half-open intervals, (*left*, *right*]. *right* == *np.inf* implies the event was right-censored.
By default, the survival function is drawn as a step function at the midpoint of each Turnbull interval.
Uses the Python *lifelines* and *matplotlib* libraries.
Uses the hpstat *turnbull* command.
:param df: Data to generate plot for
:type df: DataFrame
:param time_left: Column in *df* for the time to event, left interval endpoint (numeric)
:param time_left: Column in *df* for the time to event, left interval endpoint (numeric or timedelta)
:type time_left: str
:param time_right: Column in *df* for the time to event, right interval endpoint (numeric)
:param time_right: Column in *df* for the time to event, right interval endpoint (numeric or timedelta)
:type time_right: str
:param by: Column in *df* to stratify by (categorical)
:type by: str
:param step_loc: Proportion along the length of each Turnbull interval to step down the survival function, e.g. 0 for left bound, 1 for right bound, 0.5 for interval midpoint (numeric)
:param ci: Whether to plot confidence intervals around the survival function
:type ci: bool
:param step_loc: Proportion along the length of each Turnbull interval to step down the survival function, e.g. 0 for left bound, 1 for right bound, 0.5 for interval midpoint
:type step_loc: float
:param transform_x: Function to transform x axis by
:type transform_x: callable
@ -168,7 +175,11 @@ def turnbull(df, time_left, time_right, by=None, *, step_loc=0.5, transform_x=No
else:
df = check_nan(df[[time_left, time_right]], nan_policy)
fig, ax = plt.subplots()
# Covert timedelta to numeric
df, time_units = survtime_to_numeric(df, time_left, time_right)
if ax is None:
fig, ax = plt.subplots()
if by is not None:
# Group by independent variable
@ -176,13 +187,16 @@ def turnbull(df, time_left, time_right, by=None, *, step_loc=0.5, transform_x=No
for group in groups.groups:
subset = groups.get_group(group)
handle = plot_survfunc_turnbull(ax, subset[time_left], subset[time_right], step_loc, transform_x, transform_y)
handle = plot_survfunc_turnbull(ax, subset[time_left], subset[time_right], ci, step_loc, transform_x, transform_y)
handle.set_label('{} = {}'.format(by, group))
else:
# No grouping
plot_survfunc_turnbull(ax, df[time_left], df[time_right], step_loc, transform_x, transform_y)
plot_survfunc_turnbull(ax, df[time_left], df[time_right], ci, step_loc, transform_x, transform_y)
ax.set_xlabel('Analysis time')
if time_units:
ax.set_xlabel('{} + {} ({})'.format(time_left, time_right, time_units))
else:
ax.set_xlabel('{} + {}'.format(time_left, time_right))
ax.set_ylabel('Survival probability')
ax.set_xlim(left=0)
ax.set_ylim(0, 1)
@ -192,29 +206,39 @@ def turnbull(df, time_left, time_right, by=None, *, step_loc=0.5, transform_x=No
return fig, ax
def plot_survfunc_turnbull(ax, time_left, time_right, step_loc=0.5, transform_x=None, transform_y=None):
xpoints, ypoints = calc_survfunc_turnbull(time_left, time_right, step_loc, transform_x, transform_y)
def plot_survfunc_turnbull(ax, time_left, time_right, ci, step_loc=0.5, transform_x=None, transform_y=None):
xpoints, ypoints, ypoints0, ypoints1 = calc_survfunc_turnbull(time_left, time_right, ci, step_loc, transform_x, transform_y)
handle = ax.plot(xpoints, ypoints)[0]
if ci:
ax.fill_between(xpoints, ypoints0, ypoints1, alpha=0.3, label='_')
return handle
def calc_survfunc_turnbull(time_left, time_right, step_loc=0.5, transform_x=None, transform_y=None):
from lifelines.fitters.npmle import npmle
EPSILON = 1e-10
# TODO: Support left == right => failure was exactly observed
followup_left = time_left + EPSILON # Add epsilon to make interval half-open
followup_right = time_right
def calc_survfunc_turnbull(time_left, time_right, ci, step_loc=0.5, transform_x=None, transform_y=None):
# Estimate the survival function
#sf = lifelines.KaplanMeierFitter().fit_interval_censoring(followup_left, followup_right)
# Call lifelines.fitters.npmle.npmle directly so we can compute midpoints, etc.
sf_probs, turnbull_intervals = npmle(followup_left, followup_right)
# Prepare arguments
# TODO: Pass through other arguments
hpstat_args = [config.hpstat_path, 'turnbull', '-', '--output', 'json']
xpoints = [i.left*(1-step_loc) + i.right*step_loc for i in turnbull_intervals]
ypoints = 1 - np.cumsum(sf_probs)
# Export data to CSV
csv_buf = io.StringIO()
pd.DataFrame({'LeftTime': time_left, 'RightTime': time_right}).to_csv(csv_buf, index=False)
csv_str = csv_buf.getvalue()
# Run hpstat binary
proc = subprocess.run(hpstat_args, input=csv_str, stdout=subprocess.PIPE, stderr=None, encoding='utf-8', check=True)
raw_result = json.loads(proc.stdout)
survival_prob = np.array(raw_result['survival_prob'])
from IPython.display import clear_output
clear_output(wait=True)
xpoints = [i[0]*(1-step_loc) + i[1]*step_loc for i in raw_result['failure_intervals'] if i[1]]
ypoints = survival_prob
# Draw straight lines
# np.concatenate(...) to force starting drawing from time 0, survival 100%
@ -226,9 +250,27 @@ def calc_survfunc_turnbull(time_left, time_right, step_loc=0.5, transform_x=None
if transform_y:
ypoints = transform_y(ypoints)
return xpoints, ypoints
if ci:
zstar = -stats.norm.ppf(config.alpha/2)
survival_prob_se = np.array(raw_result['survival_prob_se'])
def survtime_to_numeric(df, time):
# Get confidence intervals
ci0 = survival_prob - zstar * survival_prob_se
ci1 = survival_prob + zstar * survival_prob_se
# Plot confidence intervals
ypoints0 = np.concatenate([[1], ci0]).repeat(2)[:-1]
ypoints1 = np.concatenate([[1], ci1]).repeat(2)[:-1]
if transform_y:
ypoints0 = transform_y(ypoints0)
ypoints1 = transform_y(ypoints1)
return xpoints, ypoints, ypoints0, ypoints1
return xpoints, ypoints, None, None
def survtime_to_numeric(df, time, time2=None):
"""
Convert pandas timedelta dtype to float64, auto-detecting the best time unit to display
@ -236,6 +278,8 @@ def survtime_to_numeric(df, time):
:type df: DataFrame
:param time: Column to check for pandas timedelta dtype
:type df: DataFrame
:param time: Second column, if any, to check for pandas timedelta dtype
:type df: DataFrame
:return: (*df*, *time_units*)
@ -243,28 +287,42 @@ def survtime_to_numeric(df, time):
* **time_units** (*str*) Human-readable description of the time unit, or *None* if not converted
"""
max_time = None
if df[time].dtype == '<m8[ns]':
df[time] = df[time].dt.total_seconds()
max_time = df[time].max()
if df[time2].dtype == '<m8[ns]':
df[time2] = df[time2].dt.total_seconds()
max_time = max(max_time or 0, df[time2].max())
if max_time is not None:
# Auto-detect best time units
if df[time].max() > 365.24*24*60*60:
df[time] = df[time] / (365.24*24*60*60)
if max_time > 365.24*24*60*60:
time_divider = 365.24*24*60*60
time_units = 'years'
elif df[time].max() > 7*24*60*60 / 12:
df[time] = df[time] / (7*24*60*60)
elif max_time > 7*24*60*60 / 12:
time_divider = 7*24*60*60
time_units = 'weeks'
elif df[time].max() > 24*60*60:
df[time] = df[time] / (24*60*60)
elif max_time > 24*60*60:
time_divider = 24*60*60
time_units = 'days'
elif df[time].max() > 60*60:
df[time] = df[time] / (60*60)
elif max_time > 60*60:
time_divider = 60*60
time_units = 'hours'
elif df[time].max() > 60:
df[time] = df[time] / 60
elif max_time > 60:
time_divider = 60
time_units = 'minutes'
else:
time_divider = 1
time_units = 'seconds'
df[time] /= time_divider
if time2:
df[time2] /= time_divider
return df, time_units
else:
return df, None