PubMedTM / pages /05_Text_Analysis.py
Array3022
2024-02-16 Release
f259e38
Raw History Blame Contribute Delete
26.3 kB
import streamlit as st
import pandas as pd
from matplotlib import pyplot as plt
import stanza
from stanza import DownloadMethod
# Load data
@st.cache_data
def load_data():
df = pd.read_pickle('medline_0_df_nlp.pkl.bz2', compression='bz2')
return df
# Notification of loading
start_container = st.empty()
if 'first' not in st.session_state:
st.session_state['first'] = True
if st.session_state['first'] == True:
start_container.text('Loading model and data... This may take two to three minutes on the initial run.')
st.session_state['first'] = False
# Load the model and data
df = load_data()
# Remove loading notification
start_container.empty()
@st.cache_resource
def load_nlp():
# Download and initialize a genia pipeline with an anatem NER model
nlp = stanza.Pipeline('en', package='genia', processors=['tokenize', 'pos', 'lemma'],
use_gpu=True, ner_batch_size=128, tokenize_batch_size=128,
download_method=DownloadMethod.REUSE_RESOURCES, logging_level='WARN')
return nlp
###### Helper Functions ######
def filter_by_year(data, year_start, year_end):
if year_start == None and year_end == None:
return data
elif year_start == None:
return data[data['DP'] <= year_end]
elif year_end == None:
return data[data['DP'] >= year_start]
else:
return data.loc[(data['DP'] >= year_start) & (data['DP'] <= year_end)]
def filter_tuples_by_entity_type(data, entity_type):
'''Filters out tuples in the NER column based on the specified entity type.'''
if entity_type == 'ALL': # Return all data
return data
data['NER'] = data['NER'].apply(lambda ner_list: [tup for tup in ner_list if tup[1] == entity_type]) # Return only tuples with the specified entity type
return data
def filter_by_country(data, country):
data = data[data['PL'] == country]
return data
def group_entity_by_year(data):
data_explode = data.explode('NER')
data_explode['NER'] = data_explode['NER'].apply(lambda x: x[0] if not type(x)==float else None) # Return none if the entity is NaN (which is seen as a float)
data_explode = data_explode.dropna(subset=['NER']) # Drop rows with None as the entity
records = zip(data_explode['DP'], data_explode['NER'])
entitiesDict = {} # {NER: [{year: count}, total_count]
for yr, name in records:
if name not in entitiesDict:
entitiesDict[name] = [{}, 0]
entityDict = entitiesDict[name][0]
if yr not in entityDict:
entityDict[yr] = 0
entityDict[yr] += 1
entitiesDict[name][1] += 1
return entitiesDict
def group_journal_by_year(data):
records = zip(data['DP'], data['JID'], data['JT'])
journalsDict = {} # {JID: [{DP: count}, total_count, JT]
for yr, jid, title in records:
if jid not in journalsDict:
journalsDict[jid] = [{}, 0, title]
journalDict = journalsDict[jid][0]
if yr not in journalDict:
journalDict[yr] = 0
journalDict[yr] += 1
journalsDict[jid][1] += 1
return journalsDict
def group_mesh_by_year(data):
data_explode = data.explode('MH')
data_explode = data_explode.dropna(subset=['MH']) # Drop rows with NaN as the entity
records = zip(data_explode['DP'], data_explode['MH'])
meshesDict = {}
for yr, word in records:
# Add mesh term if it does not exist in the dictionary
if word not in meshesDict:
meshesDict[word] = [{}, 0] # Create new entry prepared for the year count. {word: [{year: count}, total_count]}
# Update the year count
single_meshDict = meshesDict[word][0] # Access the subdictionary in each entry
if yr not in single_meshDict:
single_meshDict[yr] = 0
single_meshDict[yr] += 1
meshesDict[word][1] += 1 # Update the total count
return meshesDict
# def find_occurrence(list, target): # Finds the index of the first occurrence of an element in a list which contains tuples
# index = 0
# for element in list:
# if element[0] == target:
# return index
# index += 1
###### Collocation Functions ######
def find_collocation(word, texts, n):
# Could work on allowing user to enter list of words.
# Could remove small strings (e.g. "'s") from ngrams.
# 'America' and 'American' are different words, but 'America' and 'Americans' lemmatize to different words, which is somewhat unexpected behavior to me.
'''
Finds neighboring words/n-grams of given word in text.
Parameters:
word (str): Word to find n-grams of.
texts (list): List of strings to search for n-grams.
n (int): Number of words to include in n-gram.
Returns:
(dict): Dictionary of n-grams with word as key and count as value.
'''
# Find n-grams and create dictionary of n-grams
ngrams = {}
for text in texts:
if word in text:
i = text.index(word)
for j in range(1, n+1): # Get required number of words before and after word based on n-gram count specified
if i-j >= 0:
if text[i-j] in ngrams:
ngrams[text[i-j]] += 1
else:
ngrams[text[i-j]] = 1
try:
if text[i+j] in ngrams:
ngrams[text[i+j]] += 1
else:
ngrams[text[i+j]] = 1
except IndexError:
pass
return dict(sorted(ngrams.items(), key=lambda x:x[1], reverse=True))
# def create_plot(dictionary, word, number_of_items=10):
# '''Creates a bar chart of the frequency of the items in the dictionary'''
# fig, ax = plt.subplots()
# ax.bar(x=list(dictionary.keys())[:number_of_items], height=list(dictionary.values())[:number_of_items], align='center')
# # ax.set_xticklabels(labels=list(dictionary.keys())[:number_of_items], rotation=45)
# ax.tick_params(axis='x', rotation=45) # Rotates x-axis labels
# ax.set_xlabel('Words')
# ax.set_ylabel('Frequency')
# ax.set_title('Frequency of Words Surrounding ' + '"' + word + '"')
# return fig
def run_collocation(word, degree, dataframe=df, number_of_items=10):
nlp = load_nlp()
# Lemmatize input word
process_word = nlp(word)
lemma = process_word.sentences[0].words[0].lemma
# Find ngrams
ngrams = find_collocation(word=lemma, texts=dataframe['lemma_AB'], n=degree)
# Plot ngrams
fig, ax = plt.subplots()
ax.bar(x=list(ngrams.keys())[:number_of_items], height=list(ngrams.values())[:number_of_items], align='center')
# ax.set_xticklabels(labels=list(dictionary.keys())[:number_of_items], rotation=45)
ax.tick_params(axis='x', rotation=45) # Rotates x-axis labels
ax.set_xlabel('Words')
ax.set_ylabel('Frequency')
ax.set_title('Frequency of Words Surrounding ' + '"' + word + '"')
return fig
# return create_plot(ngrams, word)
###### Journal Analytics Functions ######
def jid_frequency(data=df, year_start=1975, year_end=2000, n=15):
'''
Returns the frequency of the top n most common journal IDs in a given year range.
'''
jid_dict = {}
added = set() # List used to keep track of all of the items which have already been added to the jid_dict
dates = data['DP'].tolist()
jid = data['JID'].tolist()
index = 0
for item in jid:
if dates[index] >= year_start and dates[index] <= year_end:
if item in added: # jid is already in the dictionary
jid_dict[item] += 1
else:
jid_dict[item] = 1
added.add(item)
index += 1
# display(jid_dict) # Used for testing and debugging
jid_df = pd.DataFrame.from_dict(jid_dict, orient='index', columns=['Frequency'])
jid_df.sort_values(by=['Frequency'], inplace=True, ascending=False)
jid_df = jid_df.head(n)
# Create matplotlib figure from dataframe
fig, ax = plt.subplots()
ax.barh(y=jid_df.index, width=jid_df['Frequency'], align='center')
ax.bar_label(ax.containers[0], fmt='%d', label_type='edge')
ax.set_xlim(right=jid_df['Frequency'].max() + jid_df['Frequency'].max()*0.08) # Prevents labels from overlapping with border in most cases; increase the number if needed
ax.set_title('Top {} Most Common JID\'s of Publications'.format(n))
ax.set_xlabel('Number of publications')
ax.set_ylabel('Journal NlmId/JID')
return fig
# jid_df.plot(kind='barh', figsize=(10, 10),
# title='Top {} Most Common JID\'s of Publications'.format(n),
# legend=True,
# xlabel='Number of publications',
# ylabel='Journal NlmId/JID',
# )
# # display(jid_df)
# return(jid_df)
def jid_frequency_by_year(data=df, year_start=1975, year_end=2000, n=15):
'''
Creates a line graph showing the frequency at which the most common n JIDs appear in the dataset for each year in given year range.
Parameters
data: the medline dataframe
n: int. The number of journals to display.
year_start: int. The start of the year range to display. Years for each data point are taken from the 'DP' column.
year_end: int. The end of the year range to display. Can be the same as year_start.
Returns
Matplotib figure object.
'''
filtered_by_year = filter_by_year(data, year_start, year_end)
journalsDict = group_journal_by_year(filtered_by_year)
sortedJournals = sorted(journalsDict.items(), key = lambda x: x[1][1], reverse = True)
topJournals = sortedJournals[0:n]
df = pd.DataFrame()
# Convert the dictionary to a Series and assign it to a new column in the DataFrame
for jid, data in topJournals:
yearly_data = data[0]
# title = data[2]
df[jid] = pd.Series(yearly_data)
# Fill NaN values with 0, in case some entities don't have data for all years
df = df.fillna(0)
# Debug
# display(df)
# Create a figure and axes
fig, ax = plt.subplots()
# Plot the data on the axes objects
for column in df.columns:
x, y = zip(*sorted(zip(df.index, df[column].values))) # Must sort the data before plotting it; https://stackoverflow.com/questions/37414916/matplotlib-plotting-in-wrong-order
ax.plot(x, y, marker='o', label=column)
# ax.plot(df.index, df[column].values)
# Set plot details
ax.set_title("Journal Frequencies by Year")
ax.set_xlabel("Year")
ax.set_ylabel("Frequency")
ax.grid(True)
ax.legend()
# ax.set_xticks([tick for tick in ax.get_xticks() if tick.is_integer()])
# ax.set_yticks([tick for tick in ax.get_yticks() if tick.is_integer()])
return fig
def pl_frequency(data=df, year_start=1975, year_end=2000, n=15):
'''
Returns the frequency of the top n most common places of publication in a given year range.
'''
pl_dict = {}
added = set() # List used to keep track of all of the items which have already been added to the ner_dict
dates = data['DP'].tolist()
pl = data['PL'].tolist()
# display(pl)
index = 0
for item in pl:
if dates[index] >= year_start and dates[index] <= year_end:
if item in added: # pl is already in the dictionary
pl_dict[item] += 1
else:
pl_dict[item] = 1
added.add(item)
index += 1
#display(pl_dict) # Used for testing and debugging
pl_df = pd.DataFrame.from_dict(
pl_dict, orient='index', columns=['Frequency'])
pl_df.sort_values(by=['Frequency'], inplace=True, ascending=False)
pl_df = pl_df.head(n)
# Create matplotlib figure from dataframe
fig, ax = plt.subplots()
ax.barh(y=pl_df.index, width=pl_df['Frequency'], align='center')
ax.bar_label(ax.containers[0], fmt='%d', label_type='edge')
ax.set_xlim(right=pl_df['Frequency'].max() + pl_df['Frequency'].max()*0.10) # Prevents labels from overlapping with border in most cases; increase the number if needed
ax.set_title('Top {} Most Common Places of Publications'.format(n))
ax.set_xlabel('Number of publications')
ax.set_ylabel('Journal NlmId/JID')
return fig
###### Named Biological Entities Functions ######
def ner_frequency(data=df, n=15, entity_toggle='ALL', year_start=1975, year_end=2000):
"""
Displays the frequency of the top n most common named entities in a graph.
Parameters
data: the medline dataframe (or a subset of it)
n: int. The number of entities to display.
entity_toggle: str. The type of entity to display. This info is stored in the tuples in the 'ner' column.
Options are 'ALL', 'ANATOMY', 'DISEASE'. If 'ALL', displays all entities.
year_start: int. The start of the year range to display.
year_end: int. The end of the year range to display. Can be the same as year_start.
"""
# Function can be slow. Could be improved by using group_entity_by_year() then counting the number of times each year appears.
ner_dict = {}
added = set() # List used to keep track of all of the items which have already been added to the ner_dict
dates = data['DP'].tolist()
ner = data['NER'].tolist()
index = 0
for item in ner:
if dates[index] >= year_start and dates[index] <= year_end:
for tup in item:
if entity_toggle == 'ALL' or entity_toggle == tup[1]:
ner = tup[0]
if ner in added: # ner is already in the dictionary
ner_dict[ner] += 1
else:
ner_dict[ner] = 1
added.add(ner)
index += 1
# display(ner_dict) # Used for testing and debugging
ner_df = pd.DataFrame.from_dict(
ner_dict, orient='index', columns=['Frequency'])
ner_df.sort_values(by=['Frequency'], inplace=True, ascending=False)
ner_df = ner_df.head(n)
fig, ax = plt.subplots()
ax.barh(y=ner_df.index, width=ner_df['Frequency'], align='center')
ax.bar_label(ax.containers[0], fmt='%d', label_type='edge')
ax.set_xlim(right=ner_df['Frequency'].max() + ner_df['Frequency'].max()*0.08) # Prevents labels from overlapping with border in most cases; increase the number if needed
if entity_toggle == 'ALL':
ax.set_title('Top {} Most Common Named Biological Entities'.format(n))
else:
ax.set_title('Top {} Most Common Named {} Entities'.format(n, entity_toggle.lower().capitalize()))
ax.set_xlabel('Number of publications')
ax.set_ylabel('Journal NlmId/JID')
return fig
def ner_frequency_by_year(data=df, n=15, entity_toggle='ALL', year_start=1975, year_end=2000) :
"""
Creates a line graph showing the frequency at which the most common n entities appear in the dataset for each year in given year range.
In other words, X axis = year, Y axis = frequency, and each line represents a different entity.
Parameters
data: the medline dataframe
n: int. The number of entities to display.
entity_toggle: str. The type of entity to display. This info is stored in the tuples in the 'NER' column.
Options are 'ALL', 'ANATOMY', 'DISEASE'. If 'ALL', displays all entities.
year_start: int. The start of the year range to display. Years for each data point are taken from the 'DP' column.
year_end: int. The end of the year range to display. Can be the same as year_start.
Returns
Matplotib figure object.
"""
filtered_by_year = filter_by_year(data, year_start, year_end)
filtered_by_entity = filter_tuples_by_entity_type(filtered_by_year, entity_toggle)
entitiesDict = group_entity_by_year(filtered_by_entity)
sortedEntities = sorted(entitiesDict.items(), key = lambda x: x[1][1], reverse = True)
topEntities = sortedEntities[0:n]
df = pd.DataFrame()
print(topEntities)
for entity, data in topEntities:
# Convert the dictionary to a Series and assign it to a new column in the DataFrame
yearly_data = data[0]
df[entity] = pd.Series(yearly_data)
# Fill NaN values with 0, in case some entities don't have data for all years
df = df.fillna(0)
print(df)
fig, ax = plt.subplots()
# Plot the data on the axes object
for column in df.columns:
x, y = zip(*sorted(zip(df.index, df[column].values))) # Must sort the data before plotting it; https://stackoverflow.com/questions/37414916/matplotlib-plotting-in-wrong-order
ax.plot(x, y, marker='o', label=column)
# Set plot details
ax.set_title("Entity Frequencies by Year")
ax.set_xlabel("Year")
ax.set_ylabel("Frequency")
ax.grid(True)
ax.legend(loc='upper left')
return fig
###### MeSH Terms Functions ######
def mh_frequency(country, globally=False, data=df, year_start=1975, year_end=2000, n=15):
'''
Returns the frequency of the top n most common MeSH terms in a given year range.
'''
mh_dict = {}
added = set() # List used to keep track of all of the items which have already been added to the mh_dict
if globally:
pass
else:
data = filter_by_country(data, country)
dates = data['DP'].tolist()
mh = data['MH'].tolist()
index = 0
for item in mh:
if dates[index] >= year_start and dates[index] <= year_end:
for i in item:
if i in added: # mh is already in the dictionary
mh_dict[i] += 1
else:
mh_dict[i] = 1
added.add(i)
index += 1
#display(mh_dict) # Used for testing and debugging
mh_df = pd.DataFrame.from_dict(
mh_dict, orient='index', columns=['Frequency'])
mh_df.sort_values(by=['Frequency'], inplace=True, ascending=False)
mh_df = mh_df.head(n)
fig, ax = plt.subplots()
ax.barh(y=mh_df.index, width=mh_df['Frequency'], align='center')
ax.bar_label(ax.containers[0], fmt='%d', label_type='edge')
ax.set_xlim(right=mh_df['Frequency'].max() + mh_df['Frequency'].max()*0.08) # Prevents labels from overlapping with border in most cases; increase the number if needed
if globally:
ax.set_title('Top {} Most Common MeSH Terms Globally'.format(n))
else:
ax.set_title('Top {} Most Common MeSH Terms in {}'.format(n, country))
ax.set_xlabel('Number of occurrences')
ax.set_ylabel('Term')
return fig
# mh_df.plot(kind='barh', figsize=(10, 10), title='Top ' +
# str(n) + ' Most Common Places of Publication', legend=False)
# #display(mh_df)
# return(mh_df)
def mh_time_series(country, globally=False, data=df, year_start=1975, year_end=2000, n=15):
"""
Creates a matplotlib line plot showing the frequency of the top n most common MeSH topics in a given country (or globally) for each year in a given range.
The MH column contains MeSH terms (controlled library of words used to categorize articles).
The PL column contains publication locations.
Parameters
country: str. The country to display.
n: int. The number of MESH topics to display.
year_start: int. The start of the year range to display. Years for each data point are taken from the 'DP' column.
year_end: int. The end of the year range to display. Can be the same as year_start.
Returns
Matplotib figure object.
# Stores the data in a dictionary with this format:
pl_mh_dict = {} # {year: {(pl, mh): (count)}} Storing the pl is redundant, but I'd like to keep it for now in case of future changes
"""
filtered_by_year = filter_by_year(data, year_start, year_end)
### TODO: Check if this is fixed
if globally:
filtered_by_country = filtered_by_year
else:
filtered_by_country = filter_by_country(filtered_by_year, country)
meshesDict = group_mesh_by_year(filtered_by_country)
sortedMeshes = sorted(meshesDict.items(), key = lambda x: x[1][1], reverse = True)
topMeshes = sortedMeshes[0:n]
df = pd.DataFrame()
for mesh, data in topMeshes:
# Convert the dictionary to a Series and assign it to a new column in the DataFrame
yearly_data = data[0]
df[mesh] = pd.Series(yearly_data)
# Fill NaN values with 0, in case some entities don't have data for all years
df = df.fillna(0)
fig, ax = plt.subplots()
# Plot the data on the axes object
for column in df.columns:
x, y = zip(*sorted(zip(df.index, df[column].values)))
ax.plot(x, y, marker='o', label=column)
# ax.plot(df.index, df[column], marker='o', label=column)
# Set plot details
if globally:
ax.set_title("MeSH Topic Frequencies by Year Globally")
else:
ax.set_title("MeSH Topic Frequencies by Year in " + country)
ax.set_xlabel("Year")
ax.set_ylabel("Frequency")
ax.grid(True)
ax.legend()
return fig
###### User Interface ######
with st.sidebar:
with st.expander('Word Collocation'):
with st.form(key='collocation_form'):
word = st.text_input(label='Enter a word to see the most common surrounding words',
placeholder='Enter a keyword')
degree = st.selectbox('Select the number of adjacent words', [1, 2, 3],
help='''If you select 1, the output will include words immediately before and after the keyword.
If you select 2, the output will include words two words before and after the keyword, etc.''')
# Could also provide an option to override the lemmatization here
submit_collocation = st.form_submit_button('Submit')
with st.expander('Journal Analytics'):
with st.form(key='journal_form'):
year_start_journal, year_end_journal = st.slider(label='Select a year range', min_value=1975, max_value=2000, value=(1975, 2000), step=1)
n_journal = st.number_input(label='Enter the number of journals to display', min_value=5, max_value=25, value=10, step=1)
submit_journal = st.form_submit_button('Submit')
with st.expander('Named Biological Entities'):
with st.form(key='nbe_form'):
entity = st.selectbox('Select an entity type', ['ALL', 'ANATOMY', 'DISEASE'],
help='''The natural language processing model that extracts these entities categorizes them as either a disease or an anatomical term. This is an artifact of the corpus that trained the model.''')
n_nbe = st.number_input(label='Enter the number of entities to display', min_value=5, max_value=25, value=10, step=1)
year_start_nbe, year_end_nbe = st.slider(label='Select a year range', min_value=1975, max_value=2000, value=(1975, 2000), step=1)
submit_nbe = st.form_submit_button('Submit')
with st.expander('MeSH Terms'):
with st.form(key='mesh_form'):
year_start_mesh, year_end_mesh = st.slider(label='Select a year range', min_value=1975, max_value=2000, value=(1975, 2000), step=1)
n_mesh = st.number_input(label='Enter the number of MeSH terms to display', min_value=5, max_value=25, value=10, step=1)
country_mesh = st.selectbox(label='Select a country to view the data related to that location',
help='The names shown here are drawn directly from data attached to the articles.',
options=sorted(df['PL'].unique().tolist()))
submit_mesh = st.form_submit_button('Submit')
if submit_collocation:
st.subheader('Collocation')
st.markdown('This program attempts to find the stem (or root form) of the word you enter and compare it against the database. This can lead to surprising results.')
collocation_plot = run_collocation(word, degree=degree, number_of_items=10)
st.pyplot(collocation_plot)
elif submit_journal:
'''
Eliminated JID metrics for now because JID is not actually a unique identifier for journals, despite claims to the contrary (https://www.nlm.nih.gov/bsd/mms/medlineelements.html#jid).
Did something go wrong in the data collection/cleaning process?
'''
# st.subheader('Journal Frequency')
# st.markdown('The NLM lists the names of journals associated with these identifiers [here](https://ftp.ncbi.nih.gov/pubmed/J_Medline.txt).')
# st.pyplot(jid_frequency(year_start=year_start_journal, year_end=year_end_journal, n=n_journal))
# st.pyplot(jid_frequency_by_year(year_start=year_start_journal, year_end=year_end_journal, n=n_journal))
# st.divider()
st.subheader('Place of Publication Frequency')
st.markdown('Note that the place of publication is derived from the location of the journal, not the location of the author(s).')
st.pyplot(pl_frequency(year_start=year_start_journal, year_end=year_end_journal, n=n_journal))
elif submit_nbe:
st.subheader('Named Biological Entities')
st.markdown('This program uses a natural language processing model to extract named biological entities from the abstracts in the database. The model categorizes these entities as either a disease or an anatomical term. This is an artifact of the corpus that trained the model.')
st.pyplot(ner_frequency(n=n_nbe, entity_toggle=entity, year_start=year_start_nbe, year_end=year_end_nbe))
st.pyplot(ner_frequency_by_year(n=n_nbe, entity_toggle=entity, year_start=year_start_nbe, year_end=year_end_nbe))
elif submit_mesh:
st.header('MeSH Term Data')
st.markdown('MeSH (Medical Subject Headings) is the NLM controlled vocabulary thesaurus used for indexing articles. For more information, visit the [MeSH homepage](https://www.nlm.nih.gov/mesh/meshhome.html).')
st.subheader('MeSH Terms by Country')
st.pyplot(mh_frequency(country=country_mesh, globally=False, data=df, year_start=year_start_mesh, year_end=year_end_mesh, n=n_mesh))
st.pyplot(mh_time_series(country=country_mesh, globally=False, data=df, year_start=year_start_mesh, year_end=year_end_mesh, n=n_mesh))
st.divider()
st.subheader('Mesh Terms Globally')
st.pyplot(mh_frequency(country=None, globally=True, data=df, year_start=year_start_mesh, year_end=year_end_mesh, n=n_mesh))
st.pyplot(mh_time_series(country=None, globally=True, data=df, year_start=year_start_mesh, year_end=year_end_mesh, n=n_mesh))